面向长跨度多方对话记忆的时序证据森林

# Temporal Evidence Forests for Long-Horizon Multi-Party Conversational Memory
Source: [https://arxiv.org/html/2610.12874](https://arxiv.org/html/2610.12874)
Junyou GuoChongyang TaoAffiliation:School of Computer Science and Engineering, Beihang University, Beijing, ChinaAffiliation:Project Lead\. Corresponding author\.Yan HeAffiliation:School of Computer Science and Engineering, Beihang University, Beijing, ChinaYong WangAffiliation:Alibaba GroupShuai MaAffiliation:School of Computer Science and Engineering, Beihang University, Beijing, China

###### Abstract

Long\-term memory has been widely adopted in LLM\-based agents to support persistent interaction and personalized assistance\. Long\-term multi\-party interactions, however, accumulate substantially more detailed information, with decisions repeatedly updated and states evolving across participants and interaction stages\. Temporal structure is therefore crucial for tracing such memories: relevant information often appears densely within short intervals, while transitions among different states are anchored by a strict temporal order\. Existing temporally aware memory frameworks primarily store time as metadata or use temporal signals to guide retrieval over preconstructed memory structures, without using temporal structure to both recover locally concentrated evidence and reconstruct its complete event progression\. To address these limitations, we proposeT\-Mem, a temporal\-structure\-based memory framework that preserves interactions and organizes query\-relevant evidence through two temporal mechanisms\. Temporal Evidence Discovery exploits the local concentration of related messages to recover overlooked evidence, while Temporal Memory Structuring models state transitions and supporting relations over time and compresses low\-information spans within the resulting forest\. Evidence Forest Linearization then serializes the structured evidence for answer generation\. Experiments on two long\-term multi\-party memory benchmarks demonstrate thatT\-Memconsistently outperforms competitive memory frameworks across LLMs, parameter settings, domains, and query types\.

## 1Introduction

An LLM agent operating over a long interaction history cannot keep every prior message in its active context\. External\-memory systems therefore preserve past interactions as passages, summaries, or structured records and retrieve a compact subset before generation\([Lewis et al\., 2020](https://arxiv.org/html/2610.12874#bib.bib1);[Packer et al\., 2023](https://arxiv.org/html/2610.12874#bib.bib3);[Xu et al\., 2026b](https://arxiv.org/html/2610.12874#bib.bib11);[Latimer et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib15);[Xu et al\., 2026a](https://arxiv.org/html/2610.12874#bib.bib16)\)\. In long\-term multi\-party interactions, however, the challenge is not only scale\. Parallel discussions are interleaved across speakers and stages, while the same issue may be repeatedly proposed, revised, rejected, or confirmed\. Semantically similar messages can therefore describe incompatible states, whereas a decisive update may share few words with the eventual query\. A useful memory context must identify the temporally coherent episode and expose the complete state evolution that determines the current answer\.

We identify two complementary temporal properties that remain poorly represented by existing memory frameworks\. The first is*temporal locality*\. Memories contributing to the same event often concentrate within a discussion episode, even when an individual message has little surface overlap with a later query\. A retrieved message can thus serve not only as evidence but also as a temporal anchor for locating its surrounding episode\. Simply increasing Top\-kkmay recover additional evidence, but it also introduces unrelated episodes and competing topics\. The question is therefore not only which messages are relevant, but also where the relevant interaction is temporally concentrated\.

The second property is*temporal evolution*, which determines how recovered evidence should be interpreted and presented\. Within an episode, messages are not interchangeable observations: a proposal may be questioned, refined, approved, implemented, or replaced, while another message may explain or support a later decision\. A similarity\-ranked list can retrieve several messages about the same subject without revealing which one represents the current state or how they form an event trajectory\. Correct answering therefore requires reconstructing the progression connecting these messages rather than treating them as independent evidence items\.

Evidence completeness alone is therefore insufficient\. When provisional proposals, later revisions, and supporting explanations are presented as an unordered set, the answer model must reconstruct both their chronology and their functional roles from scratch\. Earlier and superseded states may then compete with the final decision despite all required messages being present\. Explicitly distinguishing state\-changing updates from supporting elaborations instead converts isolated matches into an interpretable trajectory and makes the evolution leading to the answer directly observable\.

Existing agent\-memory systems improve access through graph\-based organization, persistent memory networks, multi\-granularity association, and selective consolidation\([Gutiérrez et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib6);[Rasmussen et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib10);[Chhikara et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib9);[Jiang et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib12)\)\. These approaches are generally centered on semantic association, entity structure, or pre\-query consolidation\. Although temporal information may be retained, it does not jointly serve as the primary basis for locating a relevant episode and reconstructing its internal state transitions\. Consequently, even a relevant memory set may be presented as disconnected records or transformed facts, leaving the answer model to infer which messages belong to the same event and which state supersedes another\. This limitation is particularly consequential in long\-term group conversations, where sophisticated memory systems do not consistently outperform simple retrieval\([Yang et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib13)\)\. Temporal locality and temporal evolution address complementary failure modes: the former determines where coherent evidence is concentrated, while the latter determines how the recovered messages should be interpreted\. Both are needed to recover an interaction episode and expose its complete evolution for generation\.

We introduceT\-Mem, a temporal\-structure\-aware memory framework for long\-term multi\-party interactions, built around these two objectives\.*Temporal Evidence Discovery*exploits temporal locality: Seed Retrieval first obtains anchors, whose temporal distribution identifies concentrated, topic\-specific interaction episodes\.T\-Memthen searches the original messages within these episodes to recover answer\-critical evidence that global similarity ranking may overlook\.*Temporal Memory Structuring*models temporal evolution by organizing the recovered messages into a query\-conditioned evidence forest with chronologically constrainedUpdateandExpandrelations\.Updatecaptures state progression, whereasExpandcaptures answers, explanations, and supporting evidence\. After these relations are established, low\-information spans are compressed, and Evidence Forest Linearization presents the ordered messages and their relations to the answer model\. The resulting context exposes the complete event progression rather than a flat collection of memories\.

We evaluateT\-Memon GroupMemBench\([Yang et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib13)\)and EverMemBench\([Hu et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib14)\)\.T\-Memachieves 63\.09% and 62\.68% overall accuracy on GroupMemBench with GPT\-5\.5 and Deepseek\-V4\-Flash, and 79\.46% and 77\.63% on EverMemBench, respectively\. Across both benchmarks,T\-Memconsistently outperforms flat retrieval and strong agent\-memory baselines\. Controlled comparisons support the importance of recovering temporally coherent episodes and explicitly presenting their state evolution\. Our contributions are fourfold:

- •We introduceT\-Mem, an agent\-memory framework that treats temporal structure as a core representation for retrieving and organizing evidence in long\-term multi\-party interactions\.
- •We propose Temporal Evidence Discovery, which exploits temporal locality to locate concentrated interaction episodes around retrieved anchors and recover answer\-critical messages overlooked by global similarity ranking\.
- •We develop Temporal Memory Structuring and Evidence Forest Linearization, which reconstruct chronologically grounded relations and present the complete query\-conditioned event evolution to the answer model\.
- •We conduct extensive experiments on GroupMemBench and EverMemBench, whereT\-Memachieves the highest overall accuracy across two LLM backbones and outperforms the strongest evaluated baseline by 6\.50–8\.59 percentage points\.

## 2Related Work

#### Structured and Retrieval\-Augmented Memories

Retrieval\-augmented memory systems select information from external stores to supplement the limited context of language models\. Lexical and dense retrieval provide efficient access to relevant records, while structured approaches expose broader associations and support multi\-step reasoning\. RAPTOR\([Sarthi et al\., 2024](https://arxiv.org/html/2610.12874#bib.bib4)\)recursively summarizes documents into a retrieval tree, whereas GraphRAG\([Edge et al\., 2024](https://arxiv.org/html/2610.12874#bib.bib7)\)constructs entity graphs and community summaries for query\-focused generation\. HippoRAG\([Gutiérrez et al\., 2024](https://arxiv.org/html/2610.12874#bib.bib5)\)and HippoRAG2\([Gutiérrez et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib6)\)combine knowledge graphs with associative retrieval to connect related facts\. Agent\-memory systems further transform interaction histories into reusable units: MemGPT\([Packer et al\., 2023](https://arxiv.org/html/2610.12874#bib.bib3)\)manages information through hierarchical memory tiers and context swapping, while MemGAS\([Xu et al\., 2026b](https://arxiv.org/html/2610.12874#bib.bib11)\)constructs multi\-granularity associations and adaptively selects the retrieval granularity\. These approaches broaden the information and relations available at inference time, but mainly determine which memory units are relevant\. Long\-term conversations also require determining when information applies and how it changes across interactions, motivating temporally aware memory\.

#### Temporally Aware Memories

Temporally aware memory systems incorporate temporal cues across the memory lifecycle\. A\-Mem\([Xu et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib8)\)incrementally evolves interconnected notes as new experiences arrive, allowing its memory representation to adapt over time\. Zep\([Rasmussen et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib10)\)represents facts and their validity intervals in a temporal knowledge graph for time\-aware retrieval\. Hindsight\([Latimer et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib15)\)organizes memories through temporal and entity\-aware networks and combines semantic, lexical, graph\-based, and temporal retrieval signals\. StructMem\([Xu et al\., 2026a](https://arxiv.org/html/2610.12874#bib.bib16)\)temporally anchors event\-level memories before periodically consolidating semantically related events into cross\-event representations\. PRISM\([Peng et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib17)\)performs intent\-aware retrieval and compression over a preconstructed memory graph, routing different query intents, including temporal queries, through typed relation paths\. Although these systems incorporate time at different stages, it generally serves as an auxiliary memory attribute or a signal for navigating structures organized by other relations\. Consequently, they do not explicitly organize retrieved memories into the complete progression of an event\. Furthermore, they do not exploit temporal locality to recover answer\-critical evidence concentrated around relevant interaction episodes\.

#### Memory in Long\-Term Multi\-Party Conversations

GroupMemBench\([Yang et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib13)\)and EverMemBench\([Hu et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib14)\)extend long\-term memory evaluation to multi\-party and collaborative conversations\. Unlike conventional long\-term dialogues, these conversations interleave multiple speakers, roles, topics, and concurrently evolving decisions\. Relevant evidence may therefore be distributed across participants and interaction stages, while lexically similar messages can describe different episodes or conflicting states\. Addressing this setting requires more than retrieving individually relevant memories: the system must identify when the relevant interaction occurred and organize its messages according to their temporal and evidential relationships\.T\-Memaddresses these challenges by discovering temporally localized evidence and structuring the original messages into query\-conditioned evidence forests, providing the answer model with a coherent representation of multi\-party state evolution\.

## 3Problem Formulation

Consider a group\-chat memoryℳ=\{mi\}i=1N\\mathcal\{M\}=\\\{m\_\{i\}\\\}\_\{i=1\}^\{N\}\. Each message is a tuplemi=\(xi,ti,zi\)m\_\{i\}=\(x\_\{i\},t\_\{i\},z\_\{i\}\)containing textxix\_\{i\}, timestamptit\_\{i\}, and conversation metadataziz\_\{i\}, such as author, role, channel, topic, and phase\. A queryq=\(xq,u\)q=\(x\_\{q\},u\)consists of query textxqx\_\{q\}and the requesting useruu\. For each query, the task is to derive a compact evidence context𝒮q\\mathcal\{S\}\_\{q\}fromℳ\\mathcal\{M\}and generate an answery^q=g⁡\(q,𝒮q\)\\hat\{y\}\_\{q\}=g\(q,\\mathcal\{S\}\_\{q\}\)that is supported by the conversation\. Reference answers and evidence annotations are used only for evaluation and are unavailable at inference time\.

LetL⁡\(𝒮q\)L\(\\mathcal\{S\}\_\{q\}\)denote the number of model\-input tokens in the evidence context,BBis the maximum permitted token budget\. Given an evaluation set𝒬\\mathcal\{Q\}with reference answers\{yq\}q∈𝒬\\\{y\_\{q\}\\\}\_\{q\\in\\mathcal\{Q\}\}, the objective is

max\{𝒮q\},g⁡𝔼q∼𝒬​\[J⁡\(g⁡\(q,𝒮q\),yq\)\]s\.t\.L⁡\(𝒮q\)≤B,∀q∈𝒬,\\max\_\{\\\{\\mathcal\{S\}\_\{q\}\\\},g\}\\mathbb\{E\}\_\{q\\sim\\mathcal\{Q\}\}\\\!\\left\[J\\\!\\left\(g\(q,\\mathcal\{S\}\_\{q\}\),y\_\{q\}\\right\)\\right\]\\quad\\text\{s\.t\.\}\\quad L\\\!\\left\(\\mathcal\{S\}\_\{q\}\\right\)\\leq B,\\;\\forall q\\in\\mathcal\{Q\},\(1\)whereJ⁡\(y^q,yq\)∈\{0,1\}J\(\\hat\{y\}\_\{q\},y\_\{q\}\)\\in\\\{0,1\\\}denotes the benchmark correctness judgment\. This formulation defines long\-term multi\-party memory question answering under a bounded input context without prescribing how the evidence context is obtained\.

## 4T\-Mem

### 4\.1Overview

![Refer to caption](https://arxiv.org/html/2610.12874v1/TF-Mem_Overview_FinalVersion.png)Figure 1:Overview ofT\-Mem\.Metadata\-expanded BM25 and Temporal Evidence Discovery recover candidate evidence; Temporal Memory Structuring links state evolution and support while compressing low\-information spans; and Evidence Forest Linearization presents the resulting forest\.T\-Memleaves the stored conversation unchanged and builds a compact evidence view when a query arrives\. The three stages in Figure[1](https://arxiv.org/html/2610.12874#S4.F1)discover candidate evidence, structure state changes and support, and linearize the resulting forest\. For memoryℳ=\{mi\}i=1N\\mathcal\{M\}=\\\{m\_\{i\}\\\}\_\{i=1\}^\{N\}and queryqq, the output is a forestℱq=\(𝒱q,ℰq\)\\mathcal\{F\}\_\{q\}=\(\\mathcal\{V\}\_\{q\},\\mathcal\{E\}\_\{q\}\)whose nodes are original messages\. The answer model receivesqqand a serialization of this forest\. It never sees the reference answer\. Metadata\-expanded Seed Retrieval first supplies coarse anchors, and Temporal Evidence Discovery searches the corresponding episodes to recover memories that exhibit temporal locality with these anchors\. Temporal Memory Structuring then links state\-changing and supporting relations under strict chronological constraints and compresses low\-information spans within the resulting forest\. Evidence Forest Linearization converts this structure into the answer context\.

### 4\.2Temporal Evidence Discovery

Group\-chat events often unfold within temporally concentrated episodes containing proposals, confirmations, and subsequent updates\. We first retrieve query\-aligned seed messages using a lightweight retriever \(e\.g\., BM25\), representing each message by its text together with available channel, phase, topic, author, and role metadata\. The top\-kkmessages form the coarse anchor setRk​\(q\)R\_\{k\}\(q\)\. Beyond serving as direct evidence, these seeds provide temporal anchors for locating answer\-critical messages that may have low global relevance\. We therefore use their temporal distribution to identify candidate evidence intervals within each topic and inspect the original conversations in those intervals\.

#### Topic\-wise temporal clustering\.

For each topicθ\\thetarepresented inRk​\(q\)R\_\{k\}\(q\), we define the topic\-specific seed set asAθ=\{mi∈Rk​\(q\):topici=θ\}\.A\_\{\\theta\}=\\\{m\_\{i\}\\in R\_\{k\}\(q\):\\mathrm\{topic\}\_\{i\}=\\theta\\\}\.\. We group the messages inAθA\_\{\\theta\}into temporal clusters according to the proximity of their timestamps \(e\.g\., DBSCAN or agglomerative clustering\)\. Cluster boundaries are determined by the temporal clustering procedure adopted in implementation\. For a resulting clusterCj⊆AθC\_\{j\}\\subseteq A\_\{\\theta\}with temporal spanΔ​tj\\Delta t\_\{j\}, we measure its seed density as

ρ⁡\(Cj\)=\|Cj\|max⁡\(Δ​tj,δ\),\\rho\(C\_\{j\}\)=\\frac\{\\lvert C\_\{j\}\\rvert\}\{\\max\(\\Delta t\_\{j\},\\delta\)\},\(2\)whereδ\>0\\delta\>0prevents an extremely short interval from receiving an unbounded score\. Clusters satisfyingρ⁡\(Cj\)\>μ\\rho\(C\_\{j\}\)\>\\mudefine candidate evidence intervals, which are padded to include nearby confirmations, corrections, and follow\-up updates\. The numerator favors intervals containing multiple query\-relevant seeds, while the temporal span penalizes diffuse activity\.

#### Evidence Utility Selection\.

Content similarity provides a first filter within each window\. A query\-conditioned LLM then labels the remaining messages asKey,Support, orIrrelevant\.Keymessages establish answer claims;Supportmessages provide necessary bridges or constraints\. The accepted setKjK\_\{j\}contains both labels from windowjj:

R′​\(q\)=Rk​\(q\)∪⋃jKj\.R^\{\\prime\}\(q\)=R\_\{k\}\(q\)\\cup\\bigcup\_\{j\}K\_\{j\}\.\(3\)Similarity is used for recall, but acceptance depends on evidence utility for the query\. The gate therefore does not copy a full discussion merely because its messages share keywords or a topic\.

### 4\.3Temporal Memory Structuring

T\-MemordersR′​\(q\)R^\{\\prime\}\(q\)by time and builds a temporary, query\-conditioned forest over the original messages\. A possible parent ofmjm\_\{j\}must be earlier, belong to the same topic, and respect the conversation\-stage order:

𝒫j\\displaystyle\\mathcal\{P\}\_\{j\}=\{mi∣ti<tj,topici=topicj,Oi⪯Oj\},\\displaystyle=\\\{m\_\{i\}\\mid t\_\{i\}<t\_\{j\},\\ \\text\{topic\}\_\{i\}=\\text\{topic\}\_\{j\},\\ O\_\{i\}\\preceq O\_\{j\}\\\},\(4\)Oi\\displaystyle O\_\{i\}=\(stagei,phaseOrderi,phaseProgressi\)\.\\displaystyle=\(\\text\{stage\}\_\{i\},\\text\{phaseOrder\}\_\{i\},\\text\{phaseProgress\}\_\{i\}\)\.The tupleOiO\_\{i\}collects the ordered stage and phase fields inziz\_\{i\};⪯\\preceqis lexicographic order\. A query\-conditioned classifier assigns each eligible pair one of three labels:

ei​j=Relation⁡\(q,mi,mj\)=\{Update,ifmjadvances or changes the state inmi,Expand,ifmjanswers, explains, or directly supportsmi,Root,if neither relation is supported\.e\_\{ij\}=\\operatorname\{Relation\}\(q,m\_\{i\},m\_\{j\}\)=\\begin\{cases\}\\textsc\{Update\},&\\text\{if $m\_\{j\}$ advances or changes the state in $m\_\{i\}$\},\\\\ \\textsc\{Expand\},&\\text\{if $m\_\{j\}$ answers, explains, or directly supports $m\_\{i\}$\},\\\\ \\textsc\{Root\},&\\text\{if neither relation is supported\.\}\\end\{cases\}\(5\)The highest\-confidence non\-Rootcandidate becomes the parent; otherwise,mjm\_\{j\}starts a new tree\. Temporal, topic, and stage constraints prevent backward, cross\-topic, or progression\-reversing links\. AnUpdateedge records state progression or replacement, whereas anExpandedge adds support without superseding its parent\.

The same module compresses node context when the serialized forest exceeds the answer budgetBB, without deleting nodes or changing edges\. Each leafvvis paired with its parentppand divided into spans𝒮p,v\\mathcal\{S\}\_\{p,v\}\. LetU⁡\(s\)U\(s\)denote the normalized lexical units in spanss,NNthe number of corpus messages, anddf⁡\(u\)\\operatorname\{df\}\(u\)the number containing unituu\. We define

Gq​\(s∣p\)=∑u∈U⁡\(s\)∩U⁡\(q\)\(log⁡N\+1df⁡\(u\)\+1\+1\)\+∑u∈U⁡\(s\)∖U⁡\(p\)\(log⁡N\+1df⁡\(u\)\+1\+1\),G\_\{q\}\(s\\mid p\)=\\sum\_\{u\\in U\(s\)\\cap U\(q\)\}\\left\(\\log\\frac\{N\+1\}\{\\operatorname\{df\}\(u\)\+1\}\+1\\right\)\+\\sum\_\{u\\in U\(s\)\\setminus U\(p\)\}\\left\(\\log\\frac\{N\+1\}\{\\operatorname\{df\}\(u\)\+1\}\+1\\right\),\(6\)which rewards query\-relevant content and information not already conveyed by the parent\. Its token\-normalized score is

gq​\(s∣p\)=Gq​\(s∣p\)ℓ⁡\(s\),g\_\{q\}\(s\\mid p\)=\\frac\{G\_\{q\}\(s\\mid p\)\}\{\\ell\(s\)\},\(7\)whereℓ⁡\(s\)\\ell\(s\)denotes the span length\. Spans containing query entities, timestamps, numerical values, or state changes alongUpdateedges are protected\. The remaining spans are omitted in ascending order ofgq​\(s∣p\)g\_\{q\}\(s\\mid p\)until the retained content satisfies∑s∈Zℓ⁡\(s\)≤B\\sum\_\{s\\in Z\}\\ell\(s\)\\leq B\. If necessary, compression proceeds iteratively from leaves to newly exposed parent nodes\. Every node retains its metadata, relation label, and at least one text span\.

### 4\.4Evidence Forest Linearization

We convert the structured evidence forest into a sequence that can be directly processed by the answer model\. Trees are ordered by their root timestamps, while nodes within each tree follow a preorder traversal with siblings arranged chronologically\. For each nodeviv\_\{i\}, letπi\\pi\_\{i\}denote the index of its parent andrir\_\{i\}the corresponding relation\. A root hasπi=0\\pi\_\{i\}=0andri=Rootr\_\{i\}=\\textsc\{Root\}; otherwise,ri∈\{Update,Expand\}r\_\{i\}\\in\\\{\\textsc\{Update\},\\textsc\{Expand\}\\\}\. The linearized forest is

ℒ⁡\(ℱq\)=\[\(i,πi,ri,di,ti,ai,xi\)\]i=1n,\\mathcal\{L\}\(\\mathcal\{F\}\_\{q\}\)=\\left\[\(i,\\pi\_\{i\},r\_\{i\},d\_\{i\},t\_\{i\},a\_\{i\},x\_\{i\}\)\\right\]\_\{i=1\}^\{n\},\(8\)wheredid\_\{i\},tit\_\{i\},aia\_\{i\}, andxix\_\{i\}denote the depth, timestamp, author, and retained text ofviv\_\{i\}\. Encoding each node together with its parent and relation preserves the forest structure and temporal order in a model\-readable sequence\. The answer model then generates the final response asy^=g⁡\(q,ℒ⁡\(ℱq\)\)\.\\hat\{y\}=g\\left\(q,\\mathcal\{L\}\(\\mathcal\{F\}\_\{q\}\)\\right\)\.

## 5Experiments

### 5\.1Experimental Setup

#### Datasets and Metrics\.

We evaluate on GroupMemBench\([Yang et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib13)\)and EverMemBench\([Hu et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib14)\), following their official data and evaluation protocols\. Both benchmarks test long\-term memory in multi\-party conversations\. We report answer accuracy as the primary metric\.

#### Baselines\.

We compareT\-Memwith a diverse set of strong baselines, including BM25, Qwen3\.7\-text\-embedding\([Zhang et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib2)\), A\-Mem\([Xu et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib8)\), HippoRAG2\([Gutiérrez et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib6)\), Hindsight\([Latimer et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib15)\), StructMem\([Xu et al\., 2026a](https://arxiv.org/html/2610.12874#bib.bib16)\), and MemGAS\([Xu et al\., 2026b](https://arxiv.org/html/2610.12874#bib.bib11)\)\. These baselines span conventional lexical and dense retrieval, retrieval\-augmented memory, and recent structured agent\-memory systems, providing broad coverage of representative approaches to long\-term memory\. Detailed baseline configurations are provided in Appendices[B](https://arxiv.org/html/2610.12874#A2)and[C](https://arxiv.org/html/2610.12874#A3)\.

#### Implementation Details\.

For Temporal Evidence Discovery, we use BM25 to retrievek=36k=36coarse anchors\. We instantiate the temporal clustering module with DBSCAN, clustering topic\-specific anchor timestamps using a temporal radius ofϵ=12\\epsilon=12hours andMinPts=2\\mathrm\{MinPts\}=2\. In Eq\.[2](https://arxiv.org/html/2610.12874#S4.E2), we set the span floor toδ=2\\delta=2hours and the density threshold toμ=0\.25\\mu=0\.25anchors per hour; each accepted interval is padded by 12 hours on both sides\. Messages within these intervals are ranked using qwen3\.7\-text\-embedding with 2,048\-dimensional vectors, and the Top\-100 candidates are passed to Evidence Utility Selection\. Evidence Utility Selection and Temporal Memory Structuring both use Deepseek\-V4\-Flash with thinking disabled, with temperatures of 0 and 0\.2, respectively\. The evidence context is capped atB=8192B=8192tokens\.

We separately evaluate GPT\-5\.5 and Deepseek\-V4\-Flash as both the answer and judge models\. GPT\-5\.5 uses reasoning\_effort=medium with temperature unset, whereas Deepseek\-V4\-Flash uses thinking enabled and temperature 0\.2\. The answer and judge output limits are 8,192 and 4,096 tokens, respectively\. Within each backbone, all methods use the same questions, answer and judge prompts, and decoding settings\. Additional configurations and retry policies are provided in Appendix[C](https://arxiv.org/html/2610.12874#A3)\.

### 5\.2Exp\-1: Main Evaluation

Table 1:Main results on GroupMemBench and EverMemBench\.Table[1](https://arxiv.org/html/2610.12874#S5.T1)presents the main comparison on GroupMemBench and EverMemBench\.T\-Memachieves the highest overall accuracy under both answer models on both benchmarks\. On GroupMemBench, it reaches 63\.09% with GPT\-5\.5 and 62\.68% with Deepseek\-V4\-Flash, improving over the strongest baseline by 8\.59 and 8\.05 percentage points, respectively\. The advantage remains consistent on EverMemBench, whereT\-Memexceeds the strongest completed baseline by 8\.42 and 6\.50 points\. In the Deepseek\-backed setting, StructMem is the strongest EverMemBench baseline, reaching 71\.13% overall accuracy\. These results indicate that representing temporal structure is useful beyond a particular dataset or answer model\.

The category\-level results further clarify the source of this improvement\.T\-Memperforms best on multi\-hop, knowledge\-update, and temporal questions under both backbones, where answering requires either recovering distributed evidence or distinguishing successive states\. This pattern supports the joint role of temporal evidence discovery and temporal memory structuring: the former broadens recovery within the relevant episode, while the latter preserves how supporting facts and state changes relate over time\.T\-Memdoes not dominate every isolated category—MemGAS leads GPT\-5\.5 on abstention and term ambiguity, StructMem leads Deepseek\-V4\-Flash on term ambiguity, and A\-Mem leads Deepseek\-V4\-Flash on user\-implicit reasoning and EverMemBench memory awareness\. A\-Mem’s evolving note\-level memories favor the aggregation of recurring user attributes\. StructMem’s factual–relational event decomposition and cross\-event consolidation provide multiple linked cues for disambiguating recurring terms\. Nevertheless,T\-Mem’s consistent overall lead shows a better balance across retrieval, temporal reasoning, and memory organization than either conventional retrieval or globally constructed agent memories\.

### 5\.3Exp\-2: Evidence Recovery

Figure 2:Key\-evidence recall across retrieval depths\.We test whether Temporal Evidence Discovery recovers answer\-critical messages missed by global ranking on GroupMemBench questions\. We exclude Abstention questions because the absence of answer\-bearing evidence is their defining property, making key\-evidence recall inapplicable to these queries\. We annotate canonical evidence offline by aligning gold answers with the original conversations and marking only messages that directly establish required answer claims\. Evidence\-level recall counts a unit as recovered if its canonical message or an audited, semantically equivalent alternative is retrieved; Multi\-hop units include both anchor and answer\-bearing evidence\. Gold answers and evidence labels are used only for retrospective analysis and never exposed to evaluated systems\.

Figure[2](https://arxiv.org/html/2610.12874#S5.F2)compares global retrieval baselines withT\-Mem \(Ours\)across Top\-kkbudgets\.T\-Memuses BM25 \(Expand\) hits to locate relevant temporal regions and searches within them; temporal memory structuring and evidence forest linearization are excluded\. At Top\-36, recall rises from 56\.6% with BM25 \(Expand\) and 61\.5% with HippoRAG2 to 72\.8% withT\-Mem\(\+16\.2 and \+11\.3 points, respectively\)\. Even at Top\-48, HippoRAG2 reaches only 62\.5%, versus 75\.6% forT\-Mem\. The gain therefore comes from changing*where*retrieval searches, not simply from enlarging the global candidate set\. High\-ranked messages act as temporal anchors that reopen nearby discussion episodes, allowing lower\-ranked messages that complete, revise, or resolve the target state to re\-enter the candidate set\. This directly supports the proposed temporal\-locality mechanism\. Further controlled evidence\-discovery experiments and the corresponding theoretical analysis are provided in Appendices[E](https://arxiv.org/html/2610.12874#A5)and[F](https://arxiv.org/html/2610.12874#A6), respectively\.

### 5\.4Exp\-3: Memory Organization Analysis

We isolate organization from evidence recovery by giving both conditions the same BM25 \(Expand\)\-retrieved messages at each Top\-kk\.*BM25 \(Expand\)*uses a flat chronological list, while*BM25 \(Expand\) \+ Temporal Memory Structuring*organizes the same messages as a forest and exposes UPDATE/EXPAND relations\. Temporal Evidence Discovery and the compression step of Temporal Memory Structuring are disabled, so differences reflect relation\-based structuring alone\. Answer accuracy in this analysis is evaluated with Deepseek\-V4\-Flash as both the answer and judge model\.

Figure 3:Accuracy with flat versus forest\-structured evidence\.As shown in Figure[3](https://arxiv.org/html/2610.12874#S5.F3), organization improves accuracy at every retrieval depth: the gain rises from 3\.09 points at Top\-24 to 4\.44 points at Top\-36 and 7\.60 points at Top\-48, and remains 5\.99–6\.40 points at Top\-60–72\. Flat BM25 \(Expand\) reaches 50\.74% at Top\-36 and then degrades, whereas the organized condition peaks later at 55\.70% with Top\-48\. Larger contexts therefore introduce relational interference, not just more evidence: updates, obsolete states, and supporting details become harder to distinguish in a flat list\. Explicit state\-progression and support relations make additional evidence more usable\. Thus, Exp\-2 and Exp\-3 capture complementary mechanisms: expansion recovers overlooked evidence, whereas the forest structures it for reasoning\. Appendix[E](https://arxiv.org/html/2610.12874#A5)provides a finer\-grained analysis of the topological and relational contributions to this organization effect\.

Table 2:Ablation study across answer/judge backbones\.Overall GroupMemBench accuracy \(%\) across 745 questions\. Values in parentheses are absolute changes in percentage points relative to fullT\-Memunder the same backbone\. For w/o Compression, contexts exceeding 8,192 tokens are directly truncated to the budget\.
### 5\.5Exp\-4: Ablation Study

We compare fullT\-Memwith four controlled variants over the full GroupMemBench evaluation set, holding all other settings fixed within the GPT\-5\.5 and Deepseek\-V4\-Flash backbones\. As shown in Table[2](https://arxiv.org/html/2610.12874#S5.T2), every ablation lowers accuracy under both backbones, with reductions ranging from 0\.94 to 6\.31 percentage points\. Temporal Evidence Discovery provides the largest contribution, producing gains of 6\.04 and 6\.31 points\. This result indicates that recovering evidence from the temporally localized episode addresses the principal retrieval bottleneck\. Removing temporal memory structuring yields the next\-largest losses \(4\.17 and 5\.23 points\), showing that the recovered messages remain less useful when their temporal evolution and supporting relations are not organized\. Metadata expansion and compression provide smaller but consistent improvements\. Thus, every component has a positive effect across both backbones: evidence recovery accounts for the largest improvement, temporal organization makes the recovered evidence easier to use, and compression provides a final refinement\.

Figure 4:Robustness to temporally local revision noise on Knowledge Update\.Complete Chain Recall measures recovery of all injected revisions and the final rollback; Revision w/o Rollback measures retrieval of intermediate revisions without the rollback \(lower is better\)\.
### 5\.6Exp\-5: Robustness to Context Noise

We augment the Knowledge Update subset with temporally local revision chains anchored near the original evidence\. Each chain introduces one or more intermediate revisions before a final rollback restores the original state, yielding average noise lengths of 2\.0, 2\.5, and 3\.0 messages\. The revision chains are generated with GPT\-5\.6 Sol, while answers and correctness judgments are produced with Deepseek\-V4\.1\-Flash\. Figure[4](https://arxiv.org/html/2610.12874#S5.F4)evaluates three complementary outcomes\.*Complete Chain Recall*measures whether all injected revisions and the final rollback are retrieved together, indicating that the full state trajectory is available for reasoning\.*Revision w/o Rollback*captures the failure mode in which an intermediate state is retrieved but the message that invalidates it is missed; lower is better because such partial chains favor obsolete answers\.

As the chains grow, all methods face increasing interference, whereasT\-Memmaintains stronger complete\-chain recovery and fewer revision\-only contexts\. Its accuracy decreases gradually from 47\.66% to 41\.12% and 38\.32%, rather than collapsing as more intermediate states are introduced\. This pattern shows that temporal evidence discovery more reliably reconnects related state changes and makes the rollback available alongside the revisions it supersedes\. More broadly, these results demonstrate that T\-Mem remains robust even as conversational structure becomes more complex\.

### 5\.7Exp\-6: Robustness to Metadata Availability

Figure 5:Overall GroupMemBench accuracy under restricted metadata availability\. The dashed bar marks the strongest baseline\.We examine whetherT\-Memdepends on the clean topic and phase annotations available in GroupMemBench\. We remove topic metadata from Temporal Evidence Discovery, remove both topic and phase metadata from Temporal Memory Structuring, and remove them from both modules simultaneously\. The remaining message content, timestamps, channels, authors, and roles are unchanged\. Deepseek\-V4\-Flash serves as both the answer and judge model, and all other configurations match the main evaluation\. Figure[5](https://arxiv.org/html/2610.12874#S5.F5)also includes HippoRAG2, the strongest baseline under this backbone, as a reference\.

Restricting metadata produces a gradual rather than abrupt decline\. Removing topic from evidence discovery reduces overall accuracy by only 0\.80 points, while withholding topic and phase from structuring lowers it by 1\.20 points\. Even when both modules operate without these fields,T\-Memreaches 60\.40%, remaining 5\.77 points above HippoRAG2\. The larger combined reduction indicates that topic and phase provide useful cues for separating interleaved episodes and interpreting their stages, but they are not the basis of the framework’s advantage\. Temporal proximity still localizes candidate evidence, while message content and chronological constraints retain enough information to reconstruct most state transitions\. The robustness therefore arises from the temporal mechanisms themselves rather than reliance on benchmark\-specific metadata\.

## 6Conclusion

In this paper, we proposedT\-Mem, a temporal\-structure\-based agent memory framework that treats time as a crucial representation for managing long\-term multi\-party interactions\. By exploiting temporal locality to recover evidence concentrated within relevant discussion episodes and temporal evolution to construct relations among successive states and supporting information,T\-Memeffectively addresses the challenges of locating critical evidence and reconstructing complex event evolution\. Experimental results on GroupMemBench and EverMemBench demonstrate thatT\-Memconsistently outperforms conventional retrieval methods and advanced agent memory frameworks across different LLMs and question types, while achieving favorable memory\-processing efficiency\.

### AI Use Statement

Generative AI was used for manuscript polishing and revision, as well as noise generation and data processing for a controlled experiment\. Specifically, GPT\-5\.6 Sol generated synthetic revision\-and\-rollback messages following the experiment’s data\-construction protocol\. The generation procedure is described in Appendix A\.5\. All methodological choices, experiments, numerical results, citations, and final claims were selected and verified by the authors\.

### Ethics Statement

This work follows responsible research practices and uses publicly available benchmark datasets in accordance with their intended research purposes and applicable licenses\. No private, proprietary, or personally identifiable group\-chat data were collected or used\. T\-Mem is developed to support more faithful retrieval and organization of evolving conversational memories\. Nevertheless, long\-term memory systems may expose sensitive historical information or incorrectly surface superseded decisions when deployed in real applications\. Practical deployment should therefore incorporate informed consent, access control, data minimization, retention and deletion policies, and mechanisms for auditing retrieved evidence\. All reported results are intended for research evaluation rather than automated decision\-making in high\-stakes settings\.

### Reproducibility Statement

Implementation details, metadata adaptation procedures, model configurations, evidence annotation methods, and prompt templates are documented in the appendices\.

## References

- Chhikaraet al\.\(2025\)P\. Chhikara, D\. Khant, S\. Aryan, T\. Singh, and D\. YadavMem0: building production\-ready AI agents with scalable long\-term memory\.InECAI 2025: 28th European Conference on Artificial Intelligence,Frontiers in Artificial Intelligence and Applications, Vol\.413,pp\. 2993–3000\.External Links:[Document](https://dx.doi.org/10.3233/FAIA251160),[Link](https://doi.org/10.3233/FAIA251160)Cited by:[§1](https://arxiv.org/html/2610.12874#S1.p5.1)\.
- Edgeet al\.\(2024\)D\. Edge, H\. Trinh, N\. Cheng, J\. Bradley, A\. Chao, A\. Mody, S\. Truitt, and J\. LarsonFrom local to global: a graph RAG approach to query\-focused summarization\.arXiv preprint arXiv:2404\.16130\.External Links:2404\.16130,[Document](https://dx.doi.org/10.48550/arXiv.2404.16130),[Link](https://arxiv.org/abs/2404.16130)Cited by:[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px1.p1.1)\.
- Gutiérrezet al\.\(2024\)B\. J\. Gutiérrez, Y\. Shu, Y\. Gu, M\. Yasunaga, and Y\. SuHippoRAG: neurobiologically inspired long\-term memory for large language models\.InAdvances in Neural Information Processing Systems,Vol\.37,pp\. 59532–59569\.External Links:[Document](https://dx.doi.org/10.52202/079017-1902),[Link](https://proceedings.neurips.cc/paper_files/paper/2024/file/6ddc001d07ca4f319af96a3024f6dbd1-Paper-Conference.pdf)Cited by:[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px1.p1.1)\.
- Gutiérrezet al\.\(2025\)B\. J\. Gutiérrez, Y\. Shu, W\. Qi, S\. Zhou, and Y\. SuFrom RAG to memory: non\-parametric continual learning for large language models\.InProceedings of the 42nd International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.267,pp\. 21497–21515\.External Links:[Link](https://proceedings.mlr.press/v267/gutierrez25a.html)Cited by:[Appendix B](https://arxiv.org/html/2610.12874#A2.p1.1),[§C\.2](https://arxiv.org/html/2610.12874#A3.SS2.SSS0.Px1.p2.1),[§1](https://arxiv.org/html/2610.12874#S1.p5.1),[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px1.p1.1),[§5\.1](https://arxiv.org/html/2610.12874#S5.SS1.SSS0.Px2.p1.1)\.
- Huet al\.\(2026\)C\. Hu, T\. Li, X\. Gao, H\. Chen, Y\. Bai, D\. Xu, T\. Lin, X\. Li, Y\. Han, J\. Pei, and Y\. DengEvaluating long\-horizon memory for multi\-party collaborative dialogues\.InProceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining,External Links:[Document](https://dx.doi.org/10.1145/3770855.3817589),[Link](https://doi.org/10.1145/3770855.3817589)Cited by:[§A\.2](https://arxiv.org/html/2610.12874#A1.SS2.p1.1),[Appendix B](https://arxiv.org/html/2610.12874#A2.p2.1),[§1](https://arxiv.org/html/2610.12874#S1.p7.1),[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px3.p1.1),[§5\.1](https://arxiv.org/html/2610.12874#S5.SS1.SSS0.Px1.p1.1)\.
- Jianget al\.\(2026\)D\. Jiang, Y\. Li, G\. Li, and B\. LiMAGMA: a multi\-graph based agentic memory architecture for AI agents\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),San Diego, California, United States,pp\. 36848–36865\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.1709),[Link](https://aclanthology.org/2026.acl-long.1709/)Cited by:[§1](https://arxiv.org/html/2610.12874#S1.p5.1)\.
- Latimeret al\.\(2026\)C\. Latimer, N\. Boschi, A\. Neeser, C\. Bartholomew, G\. Srivastava, X\. Wang, and N\. RamakrishnanHINDSIGHT: structured agent memory that retains, recalls, and reflects\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 3: System Demonstrations\),San Diego, California, United States,pp\. 275–285\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-demo.27),[Link](https://aclanthology.org/2026.acl-demo.27/)Cited by:[Appendix B](https://arxiv.org/html/2610.12874#A2.p1.1),[§C\.2](https://arxiv.org/html/2610.12874#A3.SS2.SSS0.Px1.p2.1),[§1](https://arxiv.org/html/2610.12874#S1.p1.1),[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px2.p1.1),[§5\.1](https://arxiv.org/html/2610.12874#S5.SS1.SSS0.Px2.p1.1)\.
- Lewiset al\.\(2020\)P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni, V\. Karpukhin, N\. Goyal, H\. Küttler, M\. Lewis, W\. Yih, T\. Rocktäschel, S\. Riedel, and D\. KielaRetrieval\-augmented generation for knowledge\-intensive NLP tasks\.InAdvances in Neural Information Processing Systems,Vol\.33,pp\. 9459–9474\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2020/file/6b493230205f780e1bc26945df7481e5-Paper.pdf)Cited by:[§1](https://arxiv.org/html/2610.12874#S1.p1.1)\.
- Packeret al\.\(2023\)C\. Packer, S\. Wooders, K\. Lin, V\. Fang, S\. G\. Patil, I\. Stoica, and J\. E\. GonzalezMemGPT: towards LLMs as operating systems\.arXiv preprint arXiv:2310\.08560\.External Links:2310\.08560,[Document](https://dx.doi.org/10.48550/arXiv.2310.08560),[Link](https://arxiv.org/abs/2310.08560)Cited by:[§1](https://arxiv.org/html/2610.12874#S1.p1.1),[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px1.p1.1)\.
- Penget al\.\(2026\)J\. Peng, Z\. Wan, W\. Liu, and Q\. SunPRISM: pareto\-efficient retrieval over intent\-aware structured memory for long\-horizon agents\.arXiv preprint arXiv:2605\.12260\.External Links:2605\.12260,[Document](https://dx.doi.org/10.48550/arXiv.2605.12260),[Link](https://arxiv.org/abs/2605.12260)Cited by:[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px2.p1.1)\.
- Rasmussenet al\.\(2025\)P\. Rasmussen, P\. Paliychuk, T\. Beauvais, J\. Ryan, and D\. ChalefZep: a temporal knowledge graph architecture for agent memory\.arXiv preprint arXiv:2501\.13956\.External Links:2501\.13956,[Document](https://dx.doi.org/10.48550/arXiv.2501.13956),[Link](https://arxiv.org/abs/2501.13956)Cited by:[§1](https://arxiv.org/html/2610.12874#S1.p5.1),[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px2.p1.1)\.
- Sarthiet al\.\(2024\)P\. Sarthi, S\. Abdullah, A\. Tuli, S\. Khanna, A\. Goldie, and C\. D\. ManningRAPTOR: recursive abstractive processing for tree\-organized retrieval\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=GN921JHCRw)Cited by:[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px1.p1.1)\.
- Xuet al\.\(2026a\)B\. Xu, Y\. Chen, J\. Fang, R\. Zhong, Y\. Yao, Y\. Zhu, L\. Du, and S\. DengStructMem: structured memory for long\-horizon behavior in LLMs\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 2: Short Papers\),San Diego, California, United States,pp\. 122–146\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-short.12),[Link](https://aclanthology.org/2026.acl-short.12/)Cited by:[Appendix B](https://arxiv.org/html/2610.12874#A2.p1.1),[§C\.2](https://arxiv.org/html/2610.12874#A3.SS2.SSS0.Px1.p2.1),[§1](https://arxiv.org/html/2610.12874#S1.p1.1),[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px2.p1.1),[§5\.1](https://arxiv.org/html/2610.12874#S5.SS1.SSS0.Px2.p1.1)\.
- Xuet al\.\(2026b\)D\. Xu, Y\. Wen, P\. Jia, Y\. Zhang, W\. Zhang, Y\. Wang, H\. Guo, R\. Tang, X\. Zhao, E\. Chen, and T\. XuFrom single to multi\-granularity: toward long\-term memory association and selection of conversational agents\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=i2yIvZARnG)Cited by:[Appendix B](https://arxiv.org/html/2610.12874#A2.p1.1),[§C\.2](https://arxiv.org/html/2610.12874#A3.SS2.SSS0.Px1.p2.1),[§1](https://arxiv.org/html/2610.12874#S1.p1.1),[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px1.p1.1),[§5\.1](https://arxiv.org/html/2610.12874#S5.SS1.SSS0.Px2.p1.1)\.
- Xuet al\.\(2025\)W\. Xu, Z\. Liang, K\. Mei, H\. Gao, J\. Tan, and Y\. ZhangA\-Mem: agentic memory for LLM agents\.InAdvances in Neural Information Processing Systems,Vol\.38, Main Conference,pp\. 17577–17604\.External Links:[Document](https://dx.doi.org/10.52202/085713-0593),[Link](https://proceedings.neurips.cc/paper_files/paper/2025/file/19909c36f51abc4856b4560aff3d36d6-Paper-Conference.pdf)Cited by:[Appendix B](https://arxiv.org/html/2610.12874#A2.p1.1),[§C\.2](https://arxiv.org/html/2610.12874#A3.SS2.SSS0.Px1.p2.1),[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px2.p1.1),[§5\.1](https://arxiv.org/html/2610.12874#S5.SS1.SSS0.Px2.p1.1)\.
- Yanget al\.\(2026\)J\. Yang, K\. Lai, X\. Wang, S\. Chang, Y\. Harari, and E\. GabrilovichGroupMemBench: benchmarking LLM agent memory in multi\-party conversations\.arXiv preprint arXiv:2605\.14498\.External Links:2605\.14498,[Document](https://dx.doi.org/10.48550/arXiv.2605.14498),[Link](https://arxiv.org/abs/2605.14498)Cited by:[§A\.1](https://arxiv.org/html/2610.12874#A1.SS1.p1.1),[Appendix B](https://arxiv.org/html/2610.12874#A2.p2.1),[§C\.2](https://arxiv.org/html/2610.12874#A3.SS2.SSS0.Px1.p1.1),[§1](https://arxiv.org/html/2610.12874#S1.p5.1),[§1](https://arxiv.org/html/2610.12874#S1.p7.1),[§2](https://arxiv.org/html/2610.12874#S2.SS0.SSS0.Px3.p1.1),[§5\.1](https://arxiv.org/html/2610.12874#S5.SS1.SSS0.Px1.p1.1)\.
- Zhanget al\.\(2025\)Y\. Zhang, M\. Li, D\. Long, X\. Zhang, H\. Lin, B\. Yang, P\. Xie, A\. Yang, D\. Liu, J\. Lin,et al\.Qwen3 embedding: advancing text embedding and reranking through foundation models\.arXiv preprint arXiv:2506\.05176\.Cited by:[§5\.1](https://arxiv.org/html/2610.12874#S5.SS1.SSS0.Px2.p1.1)\.

## Appendix ADataset and Evidence Statistics

### A\.1GroupMemBench

GroupMemBench evaluates long\-term memory over multi\-party group conversations in four domains: Finance, Technology, Healthcare, and Manufacturing\([Yang et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib13)\)\. Each domain contains 30,000 messages, giving 120,000 messages in total\. The evaluation set contains 745 questions spanning six capabilities: multi\-hop reasoning, knowledge update, temporal reasoning, user\-implicit reasoning, term ambiguity, and abstention\. Table[3](https://arxiv.org/html/2610.12874#A1.T3)reports the exact distribution used in our experiments\. We retain the official questions, gold answers, domain boundaries, and evaluation protocol\.

Table 3:GroupMemBench question distribution used in our experiments\.
### A\.2EverMemBench

EverMemBench contains long\-horizon, multi\-party and multi\-group collaborative dialogues together with 2,400 questions\([Hu et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib14)\)\. Its three first\-level evaluation dimensions are reported in the main comparison table; each dimension is itself composed of three diagnostic tasks:

- •Fine\-grained Recall: Single\-hop retrieval tests precise local grounding and entity disambiguation; Multi\-hop trajectory traces a person’s work across groups; and Temporal duration identifies event boundaries and computes elapsed time\.
- •Memory Awareness: Constraint applies implicit organizational norms to a new situation; Proactivity detects a conflict between a new instruction and a stored rule; and Update resolves rule revisions and precedence\.
- •Profile Understanding: Style reproduces an individual’s communication pattern; Skill selects actions consistent with inferred expertise; and Role answers from the speaker’s professional perspective\.

This hierarchy is important when interpreting aggregated first\-level accuracies: a single number may combine direct retrieval, cross\-message reasoning, rule application, and behavioral personalization\.

### A\.3Metadata Sources and Adaptation

Table[4](https://arxiv.org/html/2610.12874#A1.T4)reports the exact source fields used to construct the common message representation\. GroupMemBench directly provides the conversational metadata required byT\-Mem\. EverMemBench instead stores messages inside dated topic–group dialogue batches and supplies participant attributes in a separate profile table\. We join profiles by speaker name and map only source\-provided fields; in particular, EverMemBench has no native phase or reply\-parent annotation, so these fields are left empty rather than inferred\. The official question text and gold answer are read fromquestion/answerin GroupMemBench andQ/Ain EverMemBench\. Gold answers and EverMemBench’sRreferences are used only for evaluation and never enter retrieval or memory construction\. For EverMemBench, the phase component is omitted from the metadata\-expanded message representation used for Seed Retrieval; Equation \([4](https://arxiv.org/html/2610.12874#S4.E4)\) uses timestamp order within the same topic and group without introducing a synthetic phase, stage, phase order, or phase progress field\.

Table 4:Source metadata and field adaptation used for GroupMemBench and EverMemBench\. Missing metadata is left empty rather than generated\.
### A\.4Evidence Annotation and Quality for Exp\-2

Exp\-2 evaluates whether a retrieval result contains the evidence required to support the answer; it does not use answer accuracy as a proxy for retrieval\. Abstention questions are excluded because the absence of answer\-bearing evidence is their defining property\. For each of the remaining 606 questions, we align the query and gold answer to the original conversation and identify one or more required*evidence units*\. Each required message role forms one unit; when a question requires a cross\-message bridge, its anchor and answer\-bearing message are represented as separate units\. Candidate annotations are reconstructed from the original messages using a reverse\-retrieval procedure: the gold answer is used to search the original conversation, messages related to each key evidence claim are ranked by embedding similarity, and Deepseek\-V4\-Flash filters the ranked candidates for direct evidential support\. We then manually audit low\-confidence cases\. Gold answers and annotations are used only in this offline analysis and are never provided to a retriever or answer model\.

Some facts are expressed by more than one message\. We therefore annotate an alternative only when it independently establishes the same required claim; lexical overlap or topic similarity alone is insufficient\. Let𝒰q\\mathcal\{U\}\_\{q\}be the required units for questionqqand let𝒜⁡\(u\)\\mathcal\{A\}\(u\)contain the canonical message for unituutogether with its audited alternatives\. The recall contribution of an evidence unit is

rq\(u;R\)=𝕀\[R\(q\)∩𝒜\(u\)≠∅\]\.r\_\{q\}\(u;R\)=\\mathbb\{I\}\\\!\\left\[R\(q\)\\cap\\mathcal\{A\}\(u\)\\neq\\varnothing\\right\]\.\(9\)Thus, a unit is recovered by either its canonical message or one audited equivalent\. When an anchor is required, it forms a separate unit from the associated answer\-bearing evidence\. We report micro\-averaged key\-evidence recall over all questions:

Recallmicro⁡\(R\)=∑q∑u∈𝒰qrq​\(u,R\)∑q\|𝒰q\|\.\\operatorname\{Recall\}\_\{\\mathrm\{micro\}\}\(R\)=\\frac\{\\sum\_\{q\}\\sum\_\{u\\in\\mathcal\{U\}\_\{q\}\}r\_\{q\}\(u;R\)\}\{\\sum\_\{q\}\\lvert\\mathcal\{U\}\_\{q\}\\rvert\}\.\(10\)

### A\.5Noise Construction for Exp\-5

Exp\-5 measures whether a memory system can recover the final state after a sequence of temporally local revisions\. For each of the 107 Knowledge Update questions, we anchor a synthetic chain near the original answer\-bearing message\. The chain begins from the original state, introduces one to three successive revisions, and ends with a rollback that restores the original state\. The first revision is highly lexically aligned with the query, later revisions have moderate overlap, and the rollback is lexically weaker but semantically explicit\. GPT\-5\.6 Sol generates each chain from the question\-specific original state\. We require the injected event content to remain independent of and non\-duplicative with all other information in the source conversation while preserving a logically consistent progression from revision to rollback\. All injected messages follow the source metadata schema and remain close to the original evidence in time\. We construct three conditions with 2\.0, 2\.5, and 3\.0 injected messages per question on average and share each fixed realization across methods\. Hidden evaluation labels identify revisions and rollbacks but are never exposed to the retriever or answer model\.

## Appendix BComparison Baselines

We select baselines to cover distinct memory mechanisms rather than enumerate every available implementation\.BM25provides a transparent lexical reference, whileEmbeddingprovides a dense semantic retrieval baseline\. We additionally include representative agent\-memory systems:A\-Memfor dynamically linked and evolved memory notes,HippoRAG2for graph\-based associative retrieval,Hindsightfor memory organization through temporal\- and entity\-aware networks, combining semantic, lexical, graph\-based, and temporal retrieval signals,StructMemfor temporally anchored event\-level memories that are periodically consolidated into semantically related cross\-event representations, andMemGASfor multi\-granularity memory association and selection\([Xu et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib8);[Gutiérrez et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib6);[Latimer et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib15);[Xu et al\., 2026a](https://arxiv.org/html/2610.12874#bib.bib16);[Xu et al\., 2026b](https://arxiv.org/html/2610.12874#bib.bib11)\)\. Together, these systems cover sparse and dense retrieval, query\-independent memory transformation, graph organization, temporal\- and entity\-aware memory retrieval, temporally anchored event consolidation, and multi\-granularity selection\.

This selection follows the diagnostic findings of both benchmarks\. In GroupMemBench, sophisticated memory architectures do not consistently dominate simple lexical retrieval in long\-term group conversations\([Yang et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib13)\)\. EverMemBench likewise reports substantial gaps for existing memory services on multi\-party attribution, temporal revision, and implicitly relevant memories\([Hu et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib14)\)\. We therefore keep strong simple retrieval controls rather than treating system complexity as a proxy for quality\. We do not include every service or near\-duplicate memory stack evaluated by the benchmark papers\. In particular, systems whose public interface depends on a fixed user\-profile schema, omits reliable group or timestamp propagation, or changes the official evidence interface cannot be compared under the same retrieval–answer contract without confounding the memory mechanism with service\-specific behavior\. The retained set is intended to represent the principal design families while preserving a common question, answer\-model, and judging protocol\.

## Appendix CDetailed Experimental Settings

#### Model\-version clarification\.

DeepSeek\-V4\-Flash was used for the main experiments unless otherwise specified\. Following its discontinuation from the official API, subsequent experiments used DeepSeek\-V4\.1\-Flash\. The applicable model version is specified for each experiment; within each controlled comparison, all compared conditions used the same answer and judge models\.

### C\.1T\-MemDetailed Configurations

#### Memory\-processing models\.

Temporal Evidence Discovery and Temporal Memory Structuring both use Deepseek\-V4\-Flash with thinking disabled\. The former selects answer\-bearing evidence within retrieved episodes; the latter identifies directUpdateandExpandrelations under temporal\-order constraints\. Neither stage receives gold answers or annotated evidence\. Within Temporal Memory Structuring, the span\-compression procedure is deterministic and requires no additional LLM calls\.

#### Retrieval and context budgets\.

Metadata\-expanded BM25 retrieves the Top\-36 coarse anchors, indexing message content together with channel, phase, topic, author, and role\. Queries combine the asking\-user identifier with the question; no hybrid rank fusion is used\. Within each topic, temporal clustering uses a 12\-hour neighborhood and a minimum of two anchors\. Cluster density is computed as\|C\|/max⁡\(Δ​tC,2​hours\)\|C\|/\\max\(\\Delta t\_\{C\},2\\text\{ hours\}\), with threshold 0\.25 anchors per hour; accepted intervals are padded by 12 hours on either side\. The neighborhood width controls temporal continuity, while the density threshold filters weakly concentrated anchor groups\. Within each selected interval, the Top\-100 messages ranked by embedding similarity are supplied to evidence selection\. The selected evidence is merged with the coarse anchors before Temporal Memory Structuring\.

The retained evidence has a token budget ofB=8192B=8192, matching the context limit used for the baselines\. Within Temporal Memory Structuring, low\-density spans are compressed from leaves upward when this budget is exceeded, while retaining protected spans and the forest topology\. The budget concerns evidence supplied to the answer model, not the model’s output allowance; contexts already below the limit do not require budget\-driven compression\.

### C\.2Baseline Configurations

#### Models and memory construction\.

BM25 and Embedding provide direct retrieval controls without LLM\-based memory construction\. The Embedding baseline uses qwen3\.7\-text\-embedding with 2,048\-dimensional vectors and cosine similarity; BM25 uses the same metadata expansion asT\-Mem\. For A\-Mem, HippoRAG2, and Hindsight, memory construction uses Deepseek\-V4\-Flash, consistent with the ingestion backbone in GroupMemBench\([Yang et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib13)\)\. This construction model is distinct from the answer/judge backbones below\. Our StructMem adaptation partitions messages by channel, topic, and phase and splits each partition into chronological chunks of at most 24 messages\. It uses Deepseek\-V4\-Flash for factual and relational event extraction, offline memory updating, same\-channel cross\-event summarization, answering, and judging, and qwen3\.7\-text\-embedding with 2,048\-dimensional vectors for embedding operations\. Our MemGAS adaptation uses Deepseek\-V4\-Flash with thinking disabled for its LLM\-based memory stages and qwen3\.7\-text\-embedding with 2,048\-dimensional vectors for embedding operations\.

A\-Mem\([Xu et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib8)\)processes messages into notes with contextual metadata and consults five neighboring memories when establishing links and updating notes\. Its implementation refreshes the retrieval index after every 100 successful memory\-evolution updates, and query\-time retrieval returns a variable number of notes under the shared context cap\. HippoRAG2\([Gutiérrez et al\., 2025](https://arxiv.org/html/2610.12874#bib.bib6)\)performs online OpenIE with temperature 0 and a 4,096\-token extraction output limit; entity linking uses Top\-5 and final passage retrieval uses Top\-36\. Hindsight\([Latimer et al\., 2026](https://arxiv.org/html/2610.12874#bib.bib15)\)retains the conversation in domain\-specific memory banks and uses native recall withbudget=midandmax\_tokens=8192\. Its returned memory count is variable: no additional Top\-1 or fixed Top\-kkcutoff is applied\. StructMem\([Xu et al\., 2026a](https://arxiv.org/html/2610.12874#bib.bib16)\)runs in event mode and extracts both factual and relational memories\. It merges sufficiently similar updates at an offline cosine threshold of 0\.90 and forms same\-channel cross\-event summaries using a one\-hour temporal window and 15 semantic seeds\. At query time, it retrieves the Top\-60 atomic memories and Top\-5 cross\-event summaries by cosine similarity before enforcing the shared 8,192\-token context cap\. MemGAS\([Xu et al\., 2026b](https://arxiv.org/html/2610.12874#bib.bib11)\)constructs session\-, turn\-, summary\-, and keyword\-level memories, retrieves the Top\-36 sessions, and applies its multi\-granularity association and selection procedure before answer generation\.

#### Retrieval limits\.

The seven baselines share an 8,192\-token evidence\-context cap, but retain their respective retrieval units and depths, as summarized in Table[5](https://arxiv.org/html/2610.12874#A3.T5)\. A Top\-kklimit bounds the number of returned items, whereas the token cap bounds their combined length\. Consequently, short retrieval results remain unchanged; truncation is applied only when the returned context exceeds the cap\. For w/o Compression, contexts exceeding 8,192 tokens are directly truncated to the budget\.

Table 5:Baseline retrieval limits\. Context caps apply to retrieved evidence, not to generated answers or judge outputs\.

### C\.3Answer and Judge Model Settings

We evaluate two answer/judge backbones: GPT\-5\.5 and Deepseek\-V4\-Flash\. For each backbone, the same model generates answers and judges them in separate calls\. GPT\-5\.5 uses reasoning\_effort=medium; Deepseek\-V4\-Flash uses thinking enabled\. The standard output allowances are 8,192 tokens for answering and 4,096 tokens for judging, separately from the 8,192\-token evidence\-context budget\. These allowances accommodate reasoning as well as the required final response\.

Within each controlled comparison, methods use the same question set, answer/judge backbone, common answer requirements, and judging protocol\. Evidence presentation remains method\-specific: flat retrieval baselines provide passages, whereasT\-Memexposes the temporal evidence forest\. The judge receives the question, reference answer, and generated answer, accepts meaning\-equivalent paraphrases, and returnsFinal: CorrectorFinal: Incorrect\. Reference answers are never supplied to retrieval or answer generation\. For the Embedding evaluation with Deepseek\-V4\-Flash, an empty output triggers a retry with thinking disabled; completed questions are skipped rather than evaluated again\. This is an output\-recovery exception, not a change to the retrieved evidence\.

## Appendix DReproducibility

### D\.1Stage\-wise Execution

The implementation separates coarse retrieval, temporal expansion, forest organization, compression, answer generation, and judging into explicit stages\. All questions finish one stage before the next stage begins\. Each stage writes append\-only JSONL records containing the stable question key, configuration, selected message identifiers, model outputs, status, and errors\. This design allows intermediate evidence sets and forests to be inspected independently of downstream accuracy\.

### D\.2Resume and Failure Handling

Each runner reconstructs the latest record for the tuple\(domain, question\_type, question\_id\)\. Successful records are skipped on restart, while failed or incomplete records are retried\. Document embeddings, query embeddings, temporal\-window selections, and completed forest records are cached separately so that a transient model\-service or cluster failure does not require repeating earlier stages\. Structured LLM calls use bounded retries for timeouts, empty responses, and malformed JSON; persistent failures remain visible as error records rather than being silently omitted\.

### D\.3Configuration and Result Integrity

Run contracts store model names, retrieval depths, embedding dimensions, clustering parameters, prompt hashes, and output directories\. Validators check the official question counts, uniqueness of question keys, temporal ordering, single\-parent forests, and the absence of gold answers from retrieval and answer prompts\. Prompt and package manifests are hashed before execution\. Reported summaries are generated from the latest successful record per stable key, while raw JSONL histories are retained for audit\.

## Appendix EAdditional Results

### E\.1Efficiency Measurements

Table[6](https://arxiv.org/html/2610.12874#A5.T6)evaluates efficiency primarily through total token consumption, which directly reflects the amount of model processing required by each method\. Monetary cost is reported as a complementary indicator, since it also depends on the prices of the models used at different stages\. Memory\-processing tokens cover corpus construction or retrieval, whereas answer\-and\-judge tokens cover the downstream evaluation\. For Embedding, the token count includes one encoding pass over the 120,000\-message corpus and all 745 queries; because the encoder can be deployed locally, this processing does not incur an embedding API charge\. BM25 requires no model\-based memory processing\.

Table 6:Efficiency on GroupMemBench\.Token consumption is the primary efficiency measure, while total API\-equivalent cost is reported as a complementary indicator for the complete 745\-question evaluation\. Embedding tokens are counted although the encoder can be deployed locally\. The StructMem and MemGAS costs are estimated from stage\-level token usage; downstream answer and judge calls use GPT\-5\.5\-equivalent input/output pricing\.Compared with conventional retrieval\-augmented memory methods,T\-Memconsumes moderately more tokens: 31\.58M in total, compared with 14\.87M for Embedding and 5\.66M for BM25\. This additional computation supports temporal evidence recovery and structured memory organization\. Nevertheless, these memory\-processing stages use a comparatively inexpensive model, so the monetary difference remains limited:T\-Memcosts USD 44\.04, only USD 4\.40 more than Embedding and USD 5\.36 more than BM25\.

The contrast is larger for methods that use LLMs to construct or select global memory representations\. MemGAS, StructMem, HippoRAG2, A\-Mem, and Hindsight consume 41\.10M, 133\.49M, 135\.04M, 393\.19M, and 487\.35M tokens, respectively, whereasT\-Memrequires only 31\.58M\. Thus,T\-Memreduces total token consumption by 23\.2%, 76\.3%, 76\.6%, 92\.0%, and 93\.5% relative to these five baselines\. Its total cost is also lower, at USD 44\.04 versus USD 46\.84, USD 148\.67, USD 65\.50, USD 76\.70, and USD 122\.59, respectively\. These results show that query\-conditioned temporal processing avoids the high cost of globally constructing and repeatedly consolidating LLM\-derived memories\. Overall,T\-Memoffers a more efficient balance: it adds only moderate processing over traditional retrieval, yet requires far fewer tokens than existing agent memory frameworks while retaining their structured\-memory capability\.

### E\.2Evidence Discovery Analysis

Exp\-2 shows that Temporal Evidence Discovery can recover answer\-critical messages missed by global retrieval\. However, its expanded memory set may contain more messages than the initial retrieval result, so that comparison alone cannot isolate temporal locality from a larger candidate or output budget\. We therefore introduce two fixed\-budget variants\.*T\-Mem Reranked*applies Qwen3\-Reranker\-8B to the candidates recovered by Temporal Evidence Discovery and retains exactly 36 messages\.*Global Evidence Retrieval*uses the same seed results and topic partition but removes temporal localization: ifT\-Memselectsnntemporal clusters for a topic, the control instead takes the embedding Top\-100​n100ncandidates from the entire topic, applies the same Evidence Utility Selection, and reranks the survivors to Top\-36\. Both variants use qwen3\.7\-text\-embedding with 2,048\-dimensional vectors, Deepseek\-V4\-Flash for Evidence Utility Selection under the same decoding configuration, and Qwen3\-Reranker\-8B for final reranking\. All other retrieval parameters are held fixed; only the candidate search region differs\.

Figure 6:Analysis of Temporal Evidence Discovery and Temporal Memory Structuring\.Left:key\-evidence recall over 606 non\-Abstention questions\. T\-Mem Reranked and Global Evidence Retrieval both output exactly 36 messages and share the same embedding, evidence\-selection, and reranking models and configurations; the global control searches the entire topic instead of temporally localized intervals\. The horizontal axis denotes the initial retrieval depth\.Right:answer accuracy under flat, untyped\-forest, and typed temporal\-forest representations at different retrieval depths, using DeepSeek\-V4\.1\-Flash as the answer and judge model\.Figure[6](https://arxiv.org/html/2610.12874#A5.F6)shows that controlling the final context size preserves nearly all of the discovery gain: at an initial depth of 36, T\-Mem Reranked reaches 72\.3% recall, only 0\.5 percentage points below unrestrictedT\-Mem\. More importantly, T\-Mem Reranked consistently exceeds Global Evidence Retrieval even though the two controls share the same candidate capacity, final Top\-36 budget, retrieval models, and decoding settings\. At the same initial depth of 36, temporal localization raises recall from 64\.2% to 72\.3%, and the advantage persists across all tested depths\. Since the controlled variants differ only in whether dense candidate search is restricted to the query\-induced temporal intervals, the gain cannot be attributed to returning more messages, stronger models, or an additional reranking stage\. Instead, the localized intervals concentrate retrieval on the relevant episode and expose answer\-critical evidence that remains low\-ranked within the complete topic\. This provides direct controlled evidence that the benefit of Temporal Evidence Discovery arises from Temporal Locality rather than global candidate expansion\.

### E\.3Relation Semantics Analysis

Exp\-3 compares a flat list with the complete Temporal Memory Structuring module under an identical retrieved message set\. This establishes the benefit of organization, but combines the effect of the forest topology with the semantics of its relation labels\. To separate these factors, we add an untyped\-forest control\. At each Top\-kk, it preserves the parent assignments and topology of the BM25 \(Expand\) \+ Temporal Memory Structuring forest, while replacing everyUpdateandExpandlabel with a singleRelatedrelation\. All three representations use the same retrieved messages, with DeepSeek\-V4\.1\-Flash serving as both the answer and judge model\.

The right panel of Figure[6](https://arxiv.org/html/2610.12874#A5.F6)shows that both forest representations consistently outperform the flat list, while the typed temporal forest achieves the highest accuracy at every retrieval depth\. The untyped forest already yields a clear improvement, indicating that parent–child topology helps the model follow connected evidence rather than interpret each message independently\. RestoringUpdateandExpandsemantics provides a further, consistent gain; at Top\-36, for example, the typed forest reaches 58\.79% accuracy, compared with 58\.39% for the untyped forest and 54\.36% for the flat list\. The additional benefit is most pronounced when the retrieval budget is smaller, where relation labels help distinguish state\-changing revisions from supporting elaborations despite limited context\. These results indicate that the gain from Temporal Memory Structuring is driven primarily by organizing messages into an event progression, with typed relations further clarifying how each memory changes or supports that progression\.

## Appendix FTheoretical Analysis

### F\.1Reliability of Density\-Based Episode Selection

We provide an idealized analysis of the density rule in Eq\.[2](https://arxiv.org/html/2610.12874#S4.E2)\. To avoid cumbersome nested subscripts,CCbelow denotes the same topic\-specific anchor cluster written asCjC\_\{j\}in the main text\. Consider its induced candidate interval with observable durationΔC≥δ\\Delta\_\{C\}\\geq\\delta, and letAC=\|C\|A\_\{C\}=\\lvert C\\rvertbe its number of anchors\. We distinguish a relevant episode \(H1H\_\{1\}\) from background activity \(H0H\_\{0\}\)\.

#### Assumption 1 \(Anchor\-rate separation\)\.

Conditional onHhH\_\{h\}, anchor arrivals withinCCfollow a homogeneous Poisson process with rateλh\\lambda\_\{h\}, where0<λ0<λ10<\\lambda\_\{0\}<\\lambda\_\{1\}\. Hence,AC\|Hh∼Poisson⁡\(λh​ΔC\)A\_\{C\}\\mid H\_\{h\}\\sim\\operatorname\{Poisson\}\(\\lambda\_\{h\}\\Delta\_\{C\}\)\. This assumption formalizes the condition motivating Temporal Evidence Discovery: anchors are more concentrated in answer\-relevant episodes than in background intervals\.

#### Proposition 1 \(Density form of the likelihood\-ratio test\)\.

Under Assumption 1, equal class priors, and equal error costs, the likelihood\-ratio test selectsH1H\_\{1\}exactly whenAC/ΔC\>μ⋆A\_\{C\}/\\Delta\_\{C\}\>\\mu^\{\\star\}, where

μ⋆=λ1−λ0log⁡\(λ1/λ0\),λ0<μ⋆<λ1\.\\mu^\{\\star\}=\\frac\{\\lambda\_\{1\}\-\\lambda\_\{0\}\}\{\\log\(\\lambda\_\{1\}/\\lambda\_\{0\}\)\},\\qquad\\lambda\_\{0\}<\\mu^\{\\star\}<\\lambda\_\{1\}\.\(11\)Proof\.For an observed countaa, the log\-likelihood ratio is

log⁡Pr⁡\(AC=a∣H1\)Pr⁡\(AC=a∣H0\)=a​log⁡λ1λ0−\(λ1−λ0\)​ΔC\.\\log\\frac\{\\Pr\(A\_\{C\}=a\\mid H\_\{1\}\)\}\{\\Pr\(A\_\{C\}=a\\mid H\_\{0\}\)\}=a\\log\\frac\{\\lambda\_\{1\}\}\{\\lambda\_\{0\}\}\-\(\\lambda\_\{1\}\-\\lambda\_\{0\}\)\\Delta\_\{C\}\.\(12\)It is positive precisely whena/ΔC\>μ⋆a/\\Delta\_\{C\}\>\\mu^\{\\star\}\. The strict inequalities in Eq\.[11](https://arxiv.org/html/2610.12874#A6.E11)follow from the mean value theorem applied tolog⁡x\\log x\.□\\square

#### Theorem 1 \(Exponential decay of selection errors\)\.

Choose any density thresholdμ\\musatisfyingλ0<μ<λ1\\lambda\_\{0\}<\\mu<\\lambda\_\{1\}, and define

DP\(a∥b\)=alogab−a\+b\.D\_\{\\mathrm\{P\}\}\(a\\\|b\)=a\\log\\frac\{a\}\{b\}\-a\+b\.\(13\)Then the false\-positive and false\-negative probabilities obey

Pr⁡\(ACΔC≥μ\|H0\)\\displaystyle\\Pr\\\!\\left\(\\frac\{A\_\{C\}\}\{\\Delta\_\{C\}\}\\geq\\mu\\,\\middle\|\\,H\_\{0\}\\right\)≤exp\[−ΔCDP\(μ∥λ0\)\],\\displaystyle\\leq\\exp\\\!\\left\[\-\\Delta\_\{C\}D\_\{\\mathrm\{P\}\}\(\\mu\\\|\\lambda\_\{0\}\)\\right\],\(14\)Pr⁡\(ACΔC≤μ\|H1\)\\displaystyle\\Pr\\\!\\left\(\\frac\{A\_\{C\}\}\{\\Delta\_\{C\}\}\\leq\\mu\\,\\middle\|\\,H\_\{1\}\\right\)≤exp\[−ΔCDP\(μ∥λ1\)\]\.\\displaystyle\\leq\\exp\\\!\\left\[\-\\Delta\_\{C\}D\_\{\\mathrm\{P\}\}\(\\mu\\\|\\lambda\_\{1\}\)\\right\]\.\(15\)Proof\.IfA∼Poisson⁡\(λ​Δ\)A\\sim\\operatorname\{Poisson\}\(\\lambda\\Delta\), its moment\-generating function is𝔼⁡\[et​A\]=exp⁡\{λ​Δ​\(et−1\)\}\\mathbb\{E\}\[e^\{tA\}\]=\\exp\\\{\\lambda\\Delta\(e^\{t\}\-1\)\\\}\. Applying the Chernoff bound to the upper tail and minimizing overt\>0t\>0givest⋆=log⁡\(μ/λ0\)t^\{\\star\}=\\log\(\\mu/\\lambda\_\{0\}\), which yields Eq\.[14](https://arxiv.org/html/2610.12874#A6.E14)\. Applying the same argument to the lower tail witht<0t<0givest⋆=log⁡\(μ/λ1\)t^\{\\star\}=\\log\(\\mu/\\lambda\_\{1\}\)and Eq\.[15](https://arxiv.org/html/2610.12874#A6.E15)\. BecauseDP\(a∥b\)\>0D\_\{\\mathrm\{P\}\}\(a\\\|b\)\>0whenevera≠ba\\neq b, both exponents are strictly negative\.□\\square

#### Conclusion\.

Under separated anchor rates, both kinds of density\-threshold error therefore decrease exponentially with the observable interval length\. This result justifies using anchor density rather than anchor count alone\. ForΔC<δ\\Delta\_\{C\}<\\delta, the denominatormax⁡\(ΔC,δ\)\\max\(\\Delta\_\{C\},\\delta\)is a short\-interval regularizer; the exponential statement above is made only forΔC≥δ\\Delta\_\{C\}\\geq\\delta\.

#### Scope of the result\.

The guarantee is conditional on Assumption 1, an interval treated as fixed before its tested anchor count is observed, and a threshold strictly between the two anchor rates\. It establishes the reliability of the density decision under this model; it does not guarantee that the interval generator recovers every relevant episode, that real anchor arrivals are exactly homogeneous Poisson, or that selected evidence is used correctly by the answer model\. In practice, candidate intervals may be determined by a clustering algorithm or a learned selector\.

### F\.2Justification of Compression in Temporal Memory Structuring

Let𝒮\\mathcal\{S\}contain all spans considered for compression, and letpsp\_\{s\}denote the parent of the leaf containing spanss\. Letℳq\\mathcal\{M\}\_\{q\}contain all mandatory spans: protected query entities, timestamps, numerical values, state\-changing content, and one retained span per node\.

#### Assumption 2 \(Budget feasibility\)\.

The mandatory content fits the answer\-time budget:

∑s∈ℳqℓ⁡\(s\)≤B\.\\sum\_\{s\\in\\mathcal\{M\}\_\{q\}\}\\ell\(s\)\\leq B\.\(16\)Without this condition, no compression rule that preserves every protected item and every node can satisfy the budget\.

#### Lemma 1 \(Smoothed IDF as self\-information\)\.

For lexical unituu, define its smoothed empirical document\-occurrence probability by

P^​\(u\)=df⁡\(u\)\+1N\+1\.\\widehat\{P\}\(u\)=\\frac\{\\operatorname\{df\}\(u\)\+1\}\{N\+1\}\.\(17\)Its self\-information is

−log⁡P^​\(u\)=log⁡N\+1df⁡\(u\)\+1\.\-\\log\\widehat\{P\}\(u\)=\\log\\frac\{N\+1\}\{\\operatorname\{df\}\(u\)\+1\}\.\(18\)Thus, the weight in Eq\.[6](https://arxiv.org/html/2610.12874#S4.E6)is a smoothed self\-information term with a unit floor\. Rare units contribute more, while the query\-overlap and parent\-novelty sets specify which informative units are useful in context\.□\\square

Consider the budgeted span\-selection problem

max\{xs\}\\displaystyle\\max\_\{\\\{x\_\{s\}\\\}\}∑s∈𝒮xs​Gq​\(s∣ps\)\\displaystyle\\sum\_\{s\\in\\mathcal\{S\}\}x\_\{s\}G\_\{q\}\(s\\mid p\_\{s\}\)\(19\)s\.t\.\\displaystyle\\text\{s\.t\.\}∑s∈𝒮xs​ℓ​\(s\)≤B,\\displaystyle\\sum\_\{s\\in\\mathcal\{S\}\}x\_\{s\}\\ell\(s\)\\leq B,xs=1∀s∈ℳq,\\displaystyle x\_\{s\}=1\\quad\\forall s\\in\\mathcal\{M\}\_\{q\},xs∈\{0,1\}∀s∉ℳq\.\\displaystyle x\_\{s\}\\in\\\{0,1\\\}\\quad\\forall s\\notin\\mathcal\{M\}\_\{q\}\.Here,xs=1x\_\{s\}=1means that spanssis retained\. The objective preserves as much query\-conditioned information as possible under the token budget\.

#### Proposition 2 \(Optimality for the continuous relaxation\)\.

Relaxing the optional variables in Eq\.[19](https://arxiv.org/html/2610.12874#A6.E19)toxs∈\[0,1\]x\_\{s\}\\in\[0,1\]admits an optimal solution that retains optional spans in nonincreasing order of

gq​\(s∣ps\)=Gq​\(s∣ps\)ℓ⁡\(s\),g\_\{q\}\(s\\mid p\_\{s\}\)=\\frac\{G\_\{q\}\(s\\mid p\_\{s\}\)\}\{\\ell\(s\)\},\(20\)with at most one fractional boundary span after ties are resolved\.

Proof\.LetBq=B−∑s∈ℳqℓ⁡\(s\)≥0B\_\{q\}=B\-\\sum\_\{s\\in\\mathcal\{M\}\_\{q\}\}\\ell\(s\)\\geq 0be the residual budget after mandatory spans are fixed\. For the optional spans, the Lagrangian of the relaxed problem is

ℒ⁡\(x,η\)=η​Bq\+∑s∉ℳqxs​ℓ​\(s\)​\(gq​\(s∣ps\)−η\),η≥0\.\\mathcal\{L\}\(x,\\eta\)=\\eta B\_\{q\}\+\\sum\_\{s\\notin\\mathcal\{M\}\_\{q\}\}x\_\{s\}\\ell\(s\)\\bigl\(g\_\{q\}\(s\\mid p\_\{s\}\)\-\\eta\\bigr\),\\qquad\\eta\\geq 0\.\(21\)For a fixed multiplierη\\eta, separability makesxs=1x\_\{s\}=1optimal whengq​\(s∣ps\)\>ηg\_\{q\}\(s\\mid p\_\{s\}\)\>\\etaandxs=0x\_\{s\}=0optimal whengq​\(s∣ps\)<ηg\_\{q\}\(s\\mid p\_\{s\}\)<\\eta; only spans at the boundary may be fractional\. Choosingη\\etato satisfy the residual budget therefore yields the stated density ordering\.□\\square

#### Conclusion\.

Compressing nonmandatory spans from low to high information density is optimal for the continuous relaxation of the token\-budget problem\. Because the implemented procedure retains or omits whole spans, the discrete problem is a 0–1 knapsack problem; the same ordering is therefore a computationally efficient approximation rather than a claim of exact discrete optimality\. Node metadata, edges, mandatory evidence, and at least one span per node remain unchanged\.

## Appendix GExamples and Case Studies

We provide qualitative examples to complement the controlled mechanism analyses in the main paper\. These examples illustrate howT\-Mem’s two temporal mechanisms operate on individual instances and are not intended as additional causal evidence\. We first examine a case in which global retrieval identifies relevant messages but misses the message that determines the current state\.

### G\.1Recovering an Overlooked State Update

Figure[7](https://arxiv.org/html/2610.12874#A7.F7)illustrates the temporal\-locality mechanism analyzed in the evidence\-recovery experiment\. The query asks for the*current*approach to reviewing channel inconsistencies\. Although global retrieval identifies messages from the relevant discussion, both BM25 \(Top\-36\) and HippoRAG2 \(Top\-36\) miss the decisive state\-changing messageMsg\_6416\. Temporal Evidence Discovery instead uses the retrieved anchors to revisit the relevant discussion episode and recovers the overlooked update\. The recovered message explicitly supersedes the earlier “core first, assisted later” policy and specifies that the assisted channel should be reviewed together with web, email, and in\-app during the first pass\.

Temporal Evidence Discovery Recovers a Missed State UpdateInstance\.Finance / knowledge\_update\_27Question\.*What is the team’s current approach for reviewing channel inconsistencies in the customer journey?*Global retrievalBM25 \(Top\-36\): retrieves messages from the relevant discussion, butdoes not retrieveMsg\_6416\.HippoRAG2 \(Top\-36\): also missesMsg\_6416; its answer does not reflect the updated review policy\.IncorrectTemporal Evidence DiscoveryRetrieved anchors identify the relevant discussion episode\. Searching within this local region recovers the lower\-ranked but state\-changing messageMsg\_6416\.Msg\_6416— recovered key evidenceThe original “core first, assisted later” decision no longer holds because assisted matches the trigger and fallback wording, yet its owner handoff and last\-edit history diverge\.New plan:include assisted in the first pass with web/email/in\-app, and check all four together for owner drift and handoff inconsistency\.Gold answerInstead of keeping assisted as a follow\-up, include assisted in the first pass so all four channels are checked together for owner drift and handoff inconsistencies\.T\-Memanswer \(excerpt\)The team’s current approach is to lock a single unified/core onboarding journey, allow only light channel\-specific presentation differences, and have Support validateweb, email, in\-app, and assistedpaths against the spec\.CorrectFigure 7:Qualitative example of Temporal Evidence Discovery\.BM25 and HippoRAG2 at Top\-36 retrieve relevant messages but miss the state\-changing messageMsg\_6416\. Temporal locality allowsT\-Memto revisit the corresponding discussion episode, recover the overlooked update, and reconstruct the current four\-channel review policy\. Blue text highlights the answer\-critical information recovered from the local temporal episode\.The case highlights a distinction between*topic relevance*and*state completeness*\. Global retrieval can identify the correct topic while still omitting a lower\-ranked message that revises the state relevant to the query\. In this instance, the decisive information is not a new topic but a local update that changes the scope of the existing policy\. Temporal Evidence Discovery therefore recovers missing evidence by changing*where*the system searches after coarse retrieval, rather than merely increasing the global retrieval depth\. This example also shows why a current\-state question is sensitive to a single revision even when many topically relevant messages are already available\. RecoveringMsg\_6416changes the interpretation of the surrounding evidence from a staged review process to a unified first\-pass policy\. The case thus illustrates how temporal locality complements similarity ranking by restoring state\-changing evidence that would otherwise remain outside the retrieved context\.

Temporal Memory Structuring Tracks an Evolving Audit\-Retention DecisionInstance\.Manufacturing / knowledge\_update\_3Question\.*What is the current approach for ensuring compliance with audit retention requirements in the training rollout?*Flat evidence presentationThe retrieved messages contain both the earlier LMS\-export approach and later concerns about immutable audit records\. Presented as a flat list, these states are not explicitly distinguished as successive revisions\.BM25 \(Top\-36\) answer: emphasizes aligning LMS export fields and notes the concern that LMS exports alone may not satisfy Legal’s retention requirement, but does not identify the proposed append\-only logging approach or the required schema extension before the L&D build\.IncorrectTemporal Memory StructuringT\-Memlinks three messages from this discussion into an explicit sequence of state updates:Msg\_3433— earlier stateRetention fields match the LMS export schema; HRIScert\_statusandcompletion\_dateare confirmed\.Update⇓\\DownarrowMsg\_5244— challenge to the earlier approachLMS exports alone may not satisfy Legal’s retention requirement because auditors expect an immutable completion history\.Update⇓\\DownarrowMsg\_6194— current stateThe LMS export approach may not meet the immutable\-audit requirement\.New plan:add an append\-only completion log \(LRS/xAPI or an immutable audit table\) and extend the schema before the L&D build proceeds\.Gold answerInstead of relying solely on LMS export fields, implement an append\-only completion log, such as LRS/xAPI or an immutable audit table, and extend the schema before the L&D build proceeds\.T\-Memanswer \(excerpt\)The current approach is to move beyond LMS exports and add anappend\-only completion history log, either through LRS/xAPI or an immutable audit table, before the L&D build proceeds\.CorrectFigure 8:Qualitative example of Temporal Memory Structuring\.An explicitUpdatechain distinguishes the earlier LMS\-export policy, its challenge, and the revised append\-only completion plan\. Blue text marks the current update; content is condensed for readability\.
### G\.2Organizing Successive States with Temporal Relations

Figure[8](https://arxiv.org/html/2610.12874#A7.F8)illustrates how Temporal Memory Structuring represents the evolution of an audit\-retention decision\. In this example, the discussion initially relies on the LMS export schema for audit retention\. A later message questions whether these exports satisfy the immutable\-record requirement, and the latest update proposes adding an append\-only completion log and extending the schema before the L&D build proceeds\. Temporal Memory Structuring links these messages through successiveUpdaterelations to present the sequence of revisions to the answer model\.

This example illustrates how temporal relations represent successive revisions within a discussion\. The temporal evidence forest links the earlier LMS\-export approach, the subsequent challenge, and the proposed revision asMsg\_3433→\\rightarrowMsg\_5244→\\rightarrowMsg\_6194throughUpdateedges\. It complements the controlled memory\-organization analysis in the main paper by showing how the inferred relation structure is presented to the answer model\. The displayed answers alone do not establish a causal effect of structuring\. The important distinction is that all three messages remain relevant to the topic, but they do not represent equivalent states of the decision\. The explicitUpdatechain makes the direction of revision visible and prevents the earlier LMS\-only approach from being treated as co\-current with the later proposal\. This organization therefore provides the answer model with a compact representation of both decision history and state precedence\.

### G\.3Recovering a Newly Added Field Constraint

Figure[9](https://arxiv.org/html/2610.12874#A7.F9)provides a second example of how Temporal Evidence Discovery recovers a low\-ranked state update that is critical for answering a current\-state query\. The discussion contains many messages about field tagging and fallback behavior, allowing global retrieval to reach the correct topic\. However, both BM25 \(Top\-36\) and HippoRAG2 \(Top\-36\) fail to retrieveMsg\_6579, which introduces an additional fallback\-eligibility constraint\. Temporal Evidence Discovery revisits the corresponding local discussion episode and recovers this update, allowingT\-Memto reconstruct the complete two\-part tagging rule\.

Temporal Evidence Discovery Recovers a Newly Added Field ConstraintInstance\.Technology / knowledge\_update\_31Question\.*What is the current approach for tagging fields in the data integration setup?*Global retrievalBM25 \(Top\-36\): retrieves messages about field states and integration rules, butdoes not retrieveMsg\_6579\.HippoRAG2 \(Top\-36\): also missesMsg\_6579and remains focused on the earlier three\-state tagging scheme without recovering the newly added fallback\-eligibility constraint\.IncorrectTemporal Evidence DiscoveryRetrieved anchors locate the relevant field\-tagging discussion episode\. Searching within this local temporal region recovers the lower\-ranked updateMsg\_6579\.Msg\_6579— recovered key evidenceNew rule:keepstable / blocked\-for\-mapping / unstable\-but\-visiblefor display, but add a per\-fieldfallback\-eligibilityflag:safe\-copyvs\.block/hide\.Gold answerInstead of relying on the three\-state tag alone, use a two\-part rule: retain the state tag for visibility while adding an explicit fallback\-eligibility flag so that reporting can distinguish fields that may be shown with safe copy from those that should be hidden or blocked\.T\-Memanswer \(excerpt\)The current approach keeps the three display statesstable / blocked\-for\-mapping / unstable\-but\-visibleand adds a per\-fieldfallback\-eligibility flag: safe\-copy vs\. block/hide\. Reporting uses this additional flag to distinguish fields that can be safely displayed from those that should not be rendered or trusted\.CorrectFigure 9:Qualitative example of recovering a newly introduced constraint\.Global retrieval reaches the correct field\-tagging discussion but missesMsg\_6579, which changes the policy from a state tag alone to a two\-part rule containing both display state and fallback eligibility\. Temporal Evidence Discovery recovers this low\-ranked update from the corresponding local episode\. Blue text highlights the answer\-critical constraint recovered through temporal locality\.Temporal Memory Structuring Separates Revision from Execution ConstraintsInstance\.Finance / knowledge\_update\_31Question\.*What is the current approach for handling edits and ownership in the compliance policy draft?*Flat evidence presentationThe retrieved discussion contains messages about control\-impact clauses, Legal/Ops ownership, and subsequent review requirements\. In a flat presentation, however, the policy revision and the constraints that elaborate the revised policy are not explicitly distinguished\.BM25 \(Top\-36\) answer: identifies a general division in which Legal handles wording and sign\-off while Ops handles implementation\-related items, but does not recover the clause\-specific gate or its owner/backup requirements\.IncorrectTemporal Memory StructuringT\-Memorganizes the evidence into a path that distinguishes a policy update from the subsequent constraints that elaborate the new policy:Msg\_5666— preceding control stateHold the hard block only on control\-impact clauses and allow non\-control edits to continue while the Legal/Ops ownership split is finalized\.Update⇓\\DownarrowMsg\_7186— revised policyThe earlier “freeze triage first, assign ownership after full classification” procedure no longer applies\.New rule:keep non\-control edits moving, hard\-block only control\-impact clauses, and assign an owner and backup at the clause level once each item is classified\.Expand⇓\\DownarrowMsg\_9253— execution constraintsKeep@User\_10as owner and@User\_13as backup, and require a timestamp, clause\-specific change summary, and Legal re\-read before an affected clause moves forward\.Gold answerInstead of freezing the full log, use a clause\-specific gate: allow non\-control edits to continue, hard\-block only control\-impact clauses, and assign an owner and backup for each clause after classification\.T\-Memanswer \(excerpt\)The current approach is tofreeze and tightly control only risky or touched clausesrather than the full draft\. The temporal evidence forest also retains the clause\-level implementation requirements, including an owner, backup, timestamp, change summary, and Legal re\-read before affected wording proceeds\.CorrectFigure 10:Qualitative example combiningUpdateandExpandrelations\.The temporal evidence forest distinguishes the revision of the control policy from the later message that adds implementation constraints\.Updatecaptures the transition to a clause\-specific gate, whereasExpandattaches ownership and review requirements to the revised state\. Blue text highlights the current policy and the relations used to organize its supporting evidence\.This case further illustrates that topical coverage does not imply state completeness\. The retrieved context may contain extensive information about field categories while omitting a small but decisive constraint added later in the same discussion\. In this instance, the missing information does not replace the three\-state scheme; it augments it with a second decision dimension\. Recovering the local update is therefore necessary to distinguish the current two\-part rule from an incomplete description based only on the earlier state labels\. The failure mode is subtle because an answer that mentions the three display states can appear highly relevant while still missing the policy dimension that determines whether a field may be shown\. By recoveringMsg\_6579, Temporal Evidence Discovery restores the interaction between display state and fallback eligibility rather than merely adding another related message\. This makes the case a concrete example of how low\-ranked updates can change the operational meaning of an otherwise well\-covered topic\.

### G\.4Separating Policy Revision from Execution Constraints

Figure[10](https://arxiv.org/html/2610.12874#A7.F10)illustrates a more complex temporal evidence structure containing both state progression and supporting elaboration\. The discussion concerns how edits to a compliance\-policy draft should be controlled\. The central decision changes from a broad freeze\-based procedure to a clause\-specific gate, after which a later message adds execution constraints concerning ownership, timestamps, change summaries, and Legal review\. Temporal Memory Structuring separates these two functions through anUpdateedge followed by anExpandedge\.

This case illustrates that temporal organization involves more than selecting the most recent message\. Different later messages can play different roles: one may revise the governing state, while another may preserve that state and add operational detail\. Treating all later messages as interchangeable updates would obscure this distinction\. The temporal evidence forest instead representsMsg\_5666→\\rightarrowMsg\_7186as anUpdateandMsg\_7186→\\rightarrowMsg\_9253as anExpand, allowing the answer model to recover both the revised clause\-specific policy and the constraints required to execute it\. The distinction between the two edge types is important here becauseMsg\_9253does not supersede the clause\-specific gate established byMsg\_7186; it specifies how that gate should be executed\. Preserving this separation lets the temporal evidence forest encode both state precedence and supporting operational detail without collapsing them into a single notion of recency\. The resulting path therefore provides a more faithful account of how the policy evolves and becomes actionable over time\.

## Appendix HLimitations

#### Method limitations\.

First, Temporal Evidence Discovery relies on the initial Top\-kkretrieved messages as anchors for identifying temporally dense regions and triggering local evidence search\. When the initial retrieval budget is small, the anchor set may contain too few relevant messages to form a reliable cluster, or may fail to cover the relevant episode altogether\. In such cases, temporal expansion has limited ability to recover the missing evidence\. Therefore, although temporal locality can improve retrieval beyond the initial ranking, its effectiveness still depends partly on the quality and coverage of the coarse anchors\.

Second, Temporal Memory Structuring uses two broad edge types, UPDATE and EXPAND, to approximately reconstruct the evolution of an event\. This abstraction intentionally keeps the resulting memory structure simple, but the two relations may conflate semantically distinct discourse functions\. For example, questioning, clarification, justification, and elaboration may all be represented as EXPAND, even though they play different roles in the underlying interaction\. Similarly, different forms of revision or correction may be grouped under UPDATE\. Consequently, the resulting forest captures the coarse progression and supporting structure of an event rather than a complete, fine\-grained discourse representation\.

Finally, both benchmarks provide cleaner metadata and more controlled interaction histories than many deployed group\-chat systems, so generalization to noisy real\-world conversations requires further study\.

#### Analysis limitations\.

First, Exp\-2 evaluates evidence recovery against manually audited canonical evidence and identified alternatives, but the set of valid alternatives may still be incomplete\. Multi\-hop questions can admit multiple valid reasoning chains, and not every semantically equivalent message or paraphrase can be annotated\. Moreover, an LLM may answer correctly using a semantically similar message, parametric knowledge, or an unannotated paraphrase even when the canonical evidence is absent\. The alternative\-evidence audit reduces this mismatch but cannot guarantee exhaustive coverage of valid evidence\. Thus, the evidence\-recall results in Exp\-2 should be interpreted as a diagnostic estimate of recoverable answer\-supporting evidence rather than an exhaustive measure of all possible valid evidence\.

Second, the final evaluation on both benchmarks depends on LLM\-based answer generation and judging\. LLM outputs can exhibit stochastic variation, and the judge may also respond differently to semantically equivalent, partially correct, or underspecified answers\. This variability in generation and evaluation can introduce noise into the reported accuracy and may cause small variations across repeated runs\. Although we use consistent answer requirements, judging protocols, and decoding settings within controlled comparisons, these controls cannot completely eliminate model\-induced evaluation variance\.

## Appendix IAnswer and Judge Prompts

Figure[11](https://arxiv.org/html/2610.12874#A9.F11)compares the two answer interfaces\. The flat and forest variants share the same evidence\-only answering constraints and the same judge prompt\. Their only material difference is evidence presentation: the flat prompt supplies messages in retrieval\-rank order, whereas the forest prompt exposes tree boundaries andRoot/Update/Expandlabels\. Blue text marks these forest\-specific instructions\. This controlled change isolates presentation and organization from judge behavior\.

Answer and Judge Prompts: Flat vs\. Temporal Forest PresentationFlat Evidence PresentationSystem PromptYou answer questions using only the supplied retrieved messages\. Treat the message content as the sole factual source\. Do not invent facts, do not expose hidden reasoning, and do not mention the prompt, retrieval, or judging process\. Return only a direct, concise answer\.Shared User Prompt TemplateAsking user: <asking\_user\_id\>Question:<question\>Retrieved conversation passages:<serialized\_retrieved\_messages\>Answer the question using only the retrieved conversation passages above\.Temporal Forest Evidence PresentationSystem PromptYou answer questions using only the supplied retrieved messages\.The messages are organized into event\-evolution trees and serialized in preorder:•EachTREErepresents a separate event progression\.•Indentation indicates parent–child depth\.•ROOTbegins an event progression\.•UPDATEmarks a change to the parent state\.•EXPANDadds support or clarification without replacing the parent state\.•Use timestamps and these relations to reconstruct the relevant progression, followingUPDATEchains to identify the current or final state\.Treat the message content as the sole factual source\. Do not invent facts, do not expose hidden reasoning, and do not mention the prompt, retrieval, or judging process\. Return only a direct, concise answer\.Shared User Prompt TemplateAsking user: <asking\_user\_id\>Question:<question\>Retrieved conversation passages:<serialized\_retrieved\_messages\>Answer the question using only the retrieved conversation passages above\.Shared Judge System Prompt\.Judge whether the agent answer is semantically consistent with the gold answer, accepting meaning\-equivalent paraphrases\. Output onlyFinal: CorrectorFinal: Incorrect\.Shared Judge User Prompt Template\.Question: <question\>∣\\midGold Answer: <gold\_answer\>∣\\midAgent Answer: <agent\_answer\>Figure 11:Answer and judge prompts for flat and temporal\-forest evidence\. Blue text denotes instructions present only in the forest interface; the judge prompt is shared\.
## Appendix JTemporal Evidence Discovery Prompt

Evidence Utility Selection in Figure[12](https://arxiv.org/html/2610.12874#A10.F12)operates only on messages preselected from one topic–time window\. It accepts decisive evidence and strictly necessary support, while rejecting messages that are merely topically or lexically similar, obsolete, speculative, or insufficient\. The gate is not allowed to answer the question or inspect the gold answer\.

Prompt for Temporal Evidence DiscoveryStrict query\-conditioned evidence selection inside a topic–time window\.Inputs1\.One user question\.2\.Candidate messages drawn from one topic–time window after embedding preselection\.Select only1\.KEY:Messages that directly state an answer\-bearing fact or make a decisive contribution to the answer\.2\.SUPPORT:For current decisions or states, prefer the newest applicable adopted change and retain necessary scope, condition, exception, owner, deadline, or implementation status\.Reject1\.IRRELEVANT:Topic similarity, shared entities, generic background, motivation, examples, and messages that merely help locate another message\.2\.IRRELEVANT:Replaced decisions, abandoned intermediate states, and unadopted proposals when the question asks for the operative state\.3\.Weak or speculative evidence; an empty selection is preferable\.Hard constraints1\.Use only the supplied question and messages; do not access the gold answer or outside history\.2\.Do not answer the question and do not return labels for rejected messages\.Required output\{s̈elected:̈ \[\{m̈essage\_id:̈ M̈sg\_123,̈
r̈eason:̈ Ëxact answer\-bearing fact\.\}̈\]\}Figure 12:Prompt for strict, query\-conditioned evidence selection within a topic–time window\.
## Appendix KTemporal Memory Structuring Prompt

Figure[13](https://arxiv.org/html/2610.12874#A11.F13)specifies direct\-parent selection\. The query narrows relevance but cannot create a relation unsupported by the two messages\. The direct\-relation gate, deterministic ordering constraints, and single\-parent rule jointly prevent topical overlap from producing an artificially dense graph\.

Prompt for Temporal Memory StructuringQuery\-conditioned direct\-parent selection with deterministic temporal\-order gates\.Task1\.For each later child, select at most one earlier direct parent orRoot; evaluateei​j=r⁡\(q,earlier message,later message\)e\_\{ij\}=r\(q,\\text\{earlier message\},\\text\{later message\}\)\.2\.The question narrows relevance but cannot manufacture a relation that is not explicit between the two messages\.Absolute direct\-relation gate1\.Require the same concrete object, direct continuation, no hidden bridge, and counterfactual dependence\.2\.ChooseRootwhen all candidates are connected only by topic, chronology, shared vocabulary, broad relevance, or an unstated intermediate event\.Relation labels1\.Update: a direct state transition of the same decision, plan, owner, deadline, threshold, scope, implementation status, validation result, or operative rule\.2\.Expand: a direct answer, confirmation, rejection, explanation, evidence, clarification, or implementation detail for the parent’s concrete state\.3\.Root/None: no justified direct semantic parent\.Hard ordering constraints1\.Parent timestamp must be earlier; stage andphase\_ordermust be nondecreasing; within the same stage and phase,phase\_progressmust be nondecreasing\.2\.Cross\-phase edges remain eligible only when all ordering gates and the direct\-continuity test hold\.Required output\{c̈hoices:̈ \[\{c̈hild\_id:̈ M̈sg\_2,̈ p̈arent\_id:̈ M̈sg\_1,̈
r̈elation:̈ ÜPDATE\|EXPAND\|NONE,̈
c̈onfidence:̈ 0\.91, q̈uery\_relevance:̈ 0\.83\}\]\}Figure 13:Prompt for query\-conditioned Temporal Memory Structuring\.