Cognifold: Always-On Proactive Memory via Cognitive Folding

# Always-On Proactive Memory via Cognitive Folding
Source: [https://arxiv.org/html/2605.13438](https://arxiv.org/html/2605.13438)
###### Abstract

Existing agent memory remains predominantly reactive and retrieval\-based, lacking the capacity to autonomously organize experience into persistent cognitive structure\. Toward genuinely autonomous agents, we introduce Cognifold, a brain\-inspired "always\-on" agent memory designed for the next generation of proactive assistants\. CogniFold continuously folds fragmented event streams into self\-emerging cognitive structures, bootstrapping progressively higher\-level cognition from incoming events and accumulated knowledge\. We ground this by extending Complementary Learning Systems \(CLS\) theory from two layers \(hippocampus, neocortex\) to three, adding a prefrontal intent layer\. Emulating the prefrontal cortex as the locus of intentional control and decision\-making, CogniFold achieves this through graph\-topology self\-organization: cognitive structures proactively assemble under the stream, merge when semantically similar, decay when stale, relink through associative recall, and surface intents when concept\-cluster density crosses a threshold\. We evaluate structural formation using CogEval\-Bench, demonstrating that CogniFold uniquely produces memory structures that match cognitive expectations and concept emergence\. Furthermore, across 7 broad\-coverage benchmarks spanning five cognitive domains, we validate that CogniFold simultaneously performs robustly on conventional memory benchmarks\.

††footnotetext:∗Equal contribution\. 🖂 Corresponding author:duanyiquncc@gmail\.com## 1Introduction

Memory\-Augmented Agents\(Packeret al\.,[2023](https://arxiv.org/html/2605.13438#bib.bib11); Sumerset al\.,[2023](https://arxiv.org/html/2605.13438#bib.bib12)\)have empowered Large Language Models \(LLMs\) to transcend finite context constraints, enabling long\-horizon reasoning\(Shinnet al\.,[2023](https://arxiv.org/html/2605.13438#bib.bib10)\), context\-grounded personalization\(Chhikaraet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib43)\), and experience\-driven continual learning\(Majumderet al\.,[2023](https://arxiv.org/html/2605.13438#bib.bib28)\)\. However, as agents evolve from on\-demand systems into always\-on assistants, their input paradigm shifts from bounded, prompt\-driven inputs to continuously arriving, fragmented event streams\(Zacks,[2020](https://arxiv.org/html/2605.13438#bib.bib51); Kurby and Zacks,[2008](https://arxiv.org/html/2605.13438#bib.bib52)\)\. This creates an increasing demand forproactivebehaviour: an assistant that self\-organizes structure, anticipates intent, and emits goals before the user issues a query\(Einstein and McDaniel,[2005](https://arxiv.org/html/2605.13438#bib.bib4)\)\.

Yet, existing memory architectures share a common limit: their topology is fixed once formed\. Whether leveraging static knowledge graphs\(Gutiérrezet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib33),[2025](https://arxiv.org/html/2605.13438#bib.bib34)\), text\-level rewrites\(Chhikaraet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib43)\), hybrid decoupling\(Jianget al\.,[2026](https://arxiv.org/html/2605.13438#bib.bib36)\), or temporal tracking\(Rasmussenet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib68)\), memory remains agraph\-as\-product—a finished artifact to retrieve from, never a substrate thatmetabolisesunder the stream\. Consequently, agents are forced to graft proactivity on top as application\-layer machinery, such as scheduled triggers, planning loops\(Wanget al\.,[2023](https://arxiv.org/html/2605.13438#bib.bib81); Yanget al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib80)\), or periodic reflection\(Shinnet al\.,[2023](https://arxiv.org/html/2605.13438#bib.bib10); Xuet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib35)\)\. This separation creates a structural ceiling: goals can only arise from sources the application layer was explicitly designed to handle\. We argue that proactivity must instead be a property of the memory substrate—goals should emerge from the topology accumulating the conditions for them\.

![Refer to caption](https://arxiv.org/html/2605.13438v1/x2.png)Figure 1:From reactive to proactive agent memory\.Conventional agents wait for explicit user queries \(left\) or graft delayed, application\-layer triggers onto a reactive memory \(middle\)\. In contrast,CogniFold\(right\) processes unprompted, asynchronous events instantly within its memory substrate, simultaneously reactivating related dormant concepts \(e\.g\., the Vienna hotel and concert\)\.Human biological memory is evolutionarily adapted to exactly this setting: it continuously receives sensory input to autonomously encode, consolidate, forget, and surface intentions in the background\. Inspired by this, we proposeCogniFold, a proactive always\-on agent memorythat folds continuously arriving events into self\-emerging cognitive structure\.CogniFoldbootstraps in a strict sense: the graph’s current state is the interpretive context for the next event, which in turn modifies the state for future events—a self\-referential loop in which the system organises input entirely through its accumulated structure\.CogniFoldrests on two complementary perspectives\. From theneuralside, we extend Complementary Learning Systems \(CLS\) theory\(McClellandet al\.,[1995](https://arxiv.org/html/2605.13438#bib.bib1); Kumaranet al\.,[2016](https://arxiv.org/html/2605.13438#bib.bib50)\)from two layers \(hippocampus, neocortex\) to three by adding aprefrontal Intent layer; rather than being hardcoded\(Bratman,[1987](https://arxiv.org/html/2605.13438#bib.bib54)\), intents autonomously emerge once concept\-cluster density crosses a threshold\. From thecognitiveside, the graph is a substrate forconceptual bootstrapping\(Carey,[2000](https://arxiv.org/html/2605.13438#bib.bib82); Zhaoet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib83)\): recursively scaffolding higher\-level cognition from accumulated structure—a transparent, auditable form of test\-time learning, distinct from both surface\-level text rewriting \(e\.g\., A\-Mem\(Xuet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib35)\)\) and opaque gradient updates \(e\.g\., Titans\(Behrouzet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib42)\)\)\.

We conduct a two\-layer evaluation\. At thestructurallayer, we introduce CogEval\-Bench, a first\-principles evaluation framework that directly measures whether the topology formed under continuous event streams matches cognitive expectations, demonstrating thatCogniFolduniquely produces event\-grounded concepts, coherent conceptual structure, and proactive intent emergence\. At thedownstreamlayer, we evaluate across seven benchmarks spanning five cognitive domains, confirming thatCogniFoldsimultaneously performs competitively on conventional memory tasks\.

Our contributions are summarized as follows:

- •Always\-On Proactive Memory Paradigm\.We recast agent memory from a reactive retrieval target into an always\-on cognitive substrate \(Fig\.[1](https://arxiv.org/html/2605.13438#S1.F1)\), natively supporting continuous understanding and proactive anticipation\.
- •Tri\-Layered Cognitive Architecture\.We extend the two\-layer CLS framework with a prefrontal Intent layer, enabling self\-emerging intents from accumulated concepts\.
- •Continuous Topological Self\-Organization\.We identify and algorithmically resolve four intrinsic structural debts of streaming events via transparent graph\-level operations, yielding a transparent and auditable form of test\-time learning\.
- •CogEval\-Bench Evaluation Framework\.We release a structural diagnostic evaluation framework that isolates proactive emergence from retrieval accuracy\. Alongside seven established downstream benchmarks, we jointly validateCogniFold’s effectiveness in both high\-level cognitive emergence and conventional memory robustness\.

## 2CogniFold: From Neural Layers to Conceptual Bootstrapping

An always\-on agent requires a fundamentally different memory substrate\. Continuously arriving event streams demand an architecture capable of incremental, online integration\. A genuinely autonomous assistant must transition from reactive retrieval toproactive assembly—continuously capturing implicit intents and organizing relevant cognitive structures in the background\.CogniFoldgrounds this substrate in an extended Complementary Learning Systems \(CLS\) theory, formalizing memory as a dynamically evolving, typed multigraph\.

### 2\.1Tri\-Layer Substrate

![Refer to caption](https://arxiv.org/html/2605.13438v1/x3.png)Figure 2:TheCogniFoldArchitecture: Conceptual Bootstrapping via Tri\-Layered Cognitive Folding\.Extending the Complementary Learning Systems \(CLS\) framework, the memory substrate continuously metabolizes streaming events through three stages: accumulating raw episodic traces \(Hippocampal layer\), consolidating redundant patterns into semantic concepts \(Neocortical layer\), and crystallizing intents \(Prefrontal layer\)\.Human declarative memory is organized byComplementary Learning Systems\(CLS\)\(McClellandet al\.,[1995](https://arxiv.org/html/2605.13438#bib.bib1); Kumaranet al\.,[2016](https://arxiv.org/html/2605.13438#bib.bib50)\): the hippocampus rapidly encodes sparse, episode traces\(Marr,[1971](https://arxiv.org/html/2605.13438#bib.bib87); O’Reilly and McClelland,[1994](https://arxiv.org/html/2605.13438#bib.bib88); Squire,[1992](https://arxiv.org/html/2605.13438#bib.bib89)\), while the neocortex slowly distills statistical regularities into semantic representations\(Tulving and others,[1972](https://arxiv.org/html/2605.13438#bib.bib90); Pattersonet al\.,[2007](https://arxiv.org/html/2605.13438#bib.bib91)\)\. This division is dynamic: over time, long\-term storage shifts from the hippocampus to the medial prefrontal cortex \(mPFC\)\(Bontempiet al\.,[1999](https://arxiv.org/html/2605.13438#bib.bib59); Frankland and Bontempi,[2005](https://arxiv.org/html/2605.13438#bib.bib92)\)\.

Crucially, the mPFC is not a passive recipient\. It exerts top\-down control over the hippocampus via pre\-existing knowledge frameworks \(schemata\) to actively shape which hippocampal traces are retained and how they are organized\(Tseet al\.,[2007](https://arxiv.org/html/2605.13438#bib.bib60); de Sousaet al\.,[2026](https://arxiv.org/html/2605.13438#bib.bib95)\)\. This bidirectional dialogue, in which the mPFC imposes schematic frameworks to guide subsequent encoding, forms the biological substrate from which goal\-directed memory emerges\(Preston and Eichenbaum,[2013](https://arxiv.org/html/2605.13438#bib.bib61); Eichenbaum,[2017](https://arxiv.org/html/2605.13438#bib.bib93); Van Kesterenet al\.,[2012](https://arxiv.org/html/2605.13438#bib.bib94)\)\.

CogniFoldoperationalizes the three\-layer dialogue above as a typed, dynamically evolving multigraph\.Eventnodes play the hippocampal role: each input from the stream is committed verbatim and time\-stamped—an immutable episodic trace\.Conceptnodes play the neocortical role: recurrent patterns are abstracted into schemata, anchored to their constituent events through provenance edges\.Intentnodes play the prefrontal role: when concept\-level evidence converges into a coherent goal, an intent emerges and exerts top\-down influence on how subsequent events are surfaced and encoded\.

Yet, static layers are insufficient\. Structure is merely the container of cognition; the vitality of memory lies in itsmetabolism\. This brings us to the architectural dynamic at the core ofCogniFold:conceptual bootstrapping\.

### 2\.2Dynamics: Conceptual Bootstrapping

If the neuro perspective specifies the structural layers, conceptual bootstrapping\(Carey,[2000](https://arxiv.org/html/2605.13438#bib.bib82); Zhaoet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib83)\)specifies how an agent “pulls itself up by its own bootstraps” on top of them\. InCogniFold, this self\-referential dynamic unfolds throughcontinuous foldingin three stages\.

Stage 1: Accumulation\.The hippocampal layer \(Eventnodes\) ingests the raw stream verbatim\. Events initially function as cognitive*placeholders*: raw experiential fragments committed before their overarching concepts exist\.

Stage 2: Consolidation\.As events accumulate, the system detects statistical regularities across them and*consolidates*them into the neocortical layer: discrete Event nodes are folded intoConceptnodes anchored to their grounding events\.

Stage 3: Crystallization\.Concepts then act as active scaffolds for future input: incoming events are interpreted through them rather than from scratch\. When concept\-cluster density crosses a threshold, the bootstrap iterates upward—anIntentnode*crystallizes*in the prefrontal layer, providing top\-down bias for schema\-congruent encoding\. The loop closes: structure interprets experience, and experience reshapes structure\.

The synergy between neural structure and cognitive dynamic enablesCogniFoldtometaboliselike biological memory: continuously folding to eliminate redundancy \(compression\) and bootstrapping to climb levels of abstraction, sustaining cognitive agility under an always\-on event stream\.

### 2\.3Graph Formalization

Having groundedCogniFoldin neurobiological mapping \(§[2\.1](https://arxiv.org/html/2605.13438#S2.SS1)\) and cognitive dynamics \(§[2\.2](https://arxiv.org/html/2605.13438#S2.SS2)\), we now formalise the substrate as a typed directed multigraph𝒢=\(𝒱,ℛ\)\\mathcal\{G\}=\(\\mathcal\{V\},\\mathcal\{R\}\)with four node types and nine semantic edge types \(Table[1](https://arxiv.org/html/2605.13438#S2.T1)\)\.

Node types\.Event\(episodic trace\),Concept\(semantic pattern\),Intent\(crystallized goal\), andTime\(temporal anchor\)\. The first three correspond to the CLS layers;Timeis an auxiliary type connecting temporal obligations to intents viaDEADLINE\_FORedges\.

Edge ontology\.Nine typed edges encode distinct semantic relations \(Table[1](https://arxiv.org/html/2605.13438#S2.T1)\), each mapping to a specific cognitive motif\. This typed ontology constrains the LLM’s update proposals toward meaningful topology, reducing the hallucination\-driven bloat of free\-form extraction\.

Table 1:Edge types\.Each edge type maps to a specific cognitive/biological motif \(CLS Analogue column\)\. Default weights are reported in Appendix[B](https://arxiv.org/html/2605.13438#A2)\.Write/read decoupling\.The architecture decouples graph expansion from query execution\. The write path \(§[3\.1](https://arxiv.org/html/2605.13438#S3.SS1), §[3\.3](https://arxiv.org/html/2605.13438#S3.SS3)\) specifies topology\-evolution operations that run on every incoming event; the read path specifies multi\-strategy retrieval over the graph snapshot \(parameters in Appendix[B](https://arxiv.org/html/2605.13438#A2)\)\. This ensures formation and retrieval can be diagnosed independently\.

## 3Continuous Cognitive Folding

![Refer to caption](https://arxiv.org/html/2605.13438v1/x4.png)Figure 3:Continuous cognitive metabolism\.Under an asynchronous event stream, the memory substrate dynamically self\-organizes\. The graph autonomously consolidates episodic events \(Panel 3\), merges associated schemata \(Panel 4\), and crystallizes goal\-directed intents from converging concept density \(Panel 5\)\. This living topology natively supports top\-down cognitive bias \(Panel 7\), natural temporal decay \(Panel 8\), and structure\-driven proactive intervention \(Panel 9\)\.Reactive memory architectures enjoy considerable design slack: ingestion is bound to user turns, consolidation can be deferred offline, and retrieval is the only operation under latency pressure\. A proactive, always\-on agent has none of these\. Events arrive continuously and asynchronously, working memory stays bounded, and the next query may concern structure that has not yet been formed—all between user touchpoints\. The graph must therefore mutate in place under a stream that never pauses; topology must keep paying down the four structural debts—accumulation, compression, decay, completion—that any continuously evolving graph naturally accrues; and a proactive agent must assemble relevant context before being asked, which forces structural centrality, temporal recency, and usage intensity to be treated as simultaneous hard constraints rather than retrieval\-time heuristics\.

Three mechanisms purpose\-built under these pressuresoperationalise conceptual bootstrapping \(§[2\.2](https://arxiv.org/html/2605.13438#S2.SS2)\) under the stream: a proactive context\-assembly harness on the write path \(§[3\.1](https://arxiv.org/html/2605.13438#S3.SS1)\), automatic topology\-evolution operations that discharge the four debts \(§[3\.2](https://arxiv.org/html/2605.13438#S3.SS2)\), and an intent\-emergence stage that crystallizes goals from converging concept evidence \(§[3\.3](https://arxiv.org/html/2605.13438#S3.SS3)\)\.

### 3\.1Proactive Context Assembly

The working\-memory constraint is Miller’s classical capacity bound\(Miller,[1956](https://arxiv.org/html/2605.13438#bib.bib49)\)transposed to an agent: each encoding step can only reason over a tiny subset of accumulated knowledge\. This imposes apriority allocation problemon the write path: the system must select which subset of the existing graph the LLM sees when interpreting the next event\.

Priority is allocated through three signals, each anchored in a distinct cognitive memory tradition\.

Structural centrality—how embedded a node is in the cognitive graph—follows the Personalized PageRank tradition\(Gutiérrezet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib33)\); we extend it from retrieval into the write path because what the LLM sees during encoding directly shapes what it writes, making centrality a formation\-time prior rather than only a query\-time signal\.

Temporal recency—whether a trace is still fresh—follows the Ebbinghaus forgetting curve\(Ebbinghaus,[2013](https://arxiv.org/html/2605.13438#bib.bib86)\), applied to LLM memory in MemoryBank\(Zhonget al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib15)\); the same exponential kernel governs our write\-path priority so that stale schemata do not perpetually crowd out new evidence\.

Access intensity—how often a node has been re\-engaged—is a Hebbian signal\(Hebb,[2005](https://arxiv.org/html/2605.13438#bib.bib85)\): nodes that repeatedly co\-fire with the agent’s working context wire more strongly into the next context, akin to the access\-count heuristic in Mem0\(Chhikaraet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib43)\)but lifted from a passive tally to an active scoring term\.

These three signals compose linearly into a per\-node priority:

Score​\(v\)=\[α⋅PR​\(v\)\+β⋅exp⁡\(−λ⋅Δ​tv\)\+γ⋅Acc​\(v\)\]⋅U​\(v\)\\text\{Score\}\(v\)=\\bigl\[\\alpha\\cdot\\text\{PR\}\(v\)\+\\beta\\cdot\\exp\(\-\\lambda\\cdot\\Delta t\_\{v\}\)\+\\gamma\\cdot\\text\{Acc\}\(v\)\\bigr\]\\cdot U\(v\)\(1\)whereU​\(v\)≥1U\(v\)\\geq 1is a deadline\-driven urgency multiplier from connectedTimenodes; weights are reported in Appendix[B](https://arxiv.org/html/2605.13438#A2)\.

The resulting scores define aproactive context window: structurally central, temporally fresh, and frequently used knowledge surfaces before the next event is interpreted—rather than waiting for a later query to reveal what should have mattered\. Selected nodes are partitioned into Immediate, Working, and Background tiers \(proportions in Appendix[B](https://arxiv.org/html/2605.13438#A2)\), forcing the LLM to processete\_\{t\}within a layered subgraph rather than in isolation\.

Given the assembled context, an LLM central executive emits anUpdatePlan: a sequence of atomic operations \(ADD\_NODE,ADD\_EDGE,UPDATE\_NODE,MERGE\_NODES,REMOVE\_NODE\), each carrying natural\-language reasoning andgrounded\_inprovenance\. The executor validates and applies the plan atomically with snapshot\-based rollback, and near\-duplicate concepts \(above a title\-similarity threshold; see Appendix[B](https://arxiv.org/html/2605.13438#A2)\) are silently converted to reinforcement updates to prevent bloat from redundant events\.

### 3\.2Four Structural Debts

Continuous event arrival is not a neutral inflow: by the nature of the input, a memory graph accumulates four kinds of structural debt over time\. These are not design choices but mandatory state\-change operations imposed by the stream; any always\-on memory that fails to address any one of them degrades along the corresponding axis\.CogniFoldaddresses all four as automatic graph\-level operations, executed without per\-step LLM supervision in a consolidation pass inspired by sleep\-dependent consolidation\(Stickgold and Walker,[2007](https://arxiv.org/html/2605.13438#bib.bib53)\)\.

1. 1\.Accumulation—persistent patterns must strengthen; one\-off noise must not\. Without it, a one\-off event and a recurring concept reach equivalent PageRank\. Our operation: when a new event corroborates an existing concept, the system creates aREINFORCESedge rather than a duplicate node, boosting that concept’s in\-degree and PageRank—implementing Bartlett’s schema assimilation\(Bartlett,[1995](https://arxiv.org/html/2605.13438#bib.bib7)\)as a graph operation\.
2. 2\.Compression—redundant fragments must fold\. Without it, graph size grows with\|events\|\|\\text\{events\}\|, PageRank diffuses across duplicates, and evidence that should aggregate stays fragmented\. Our operation: when two concept nodes exceed a semantic\-similarity threshold \(Appendix[B](https://arxiv.org/html/2605.13438#A2)\), the executor automatically merges them \(MERGE\_NODES\); the higher\-access node absorbs all edges\. This implements schema unitization\(Gilboa and Marlatte,[2017](https://arxiv.org/html/2605.13438#bib.bib6)\)and physically shortens graph\-theoretic distances—multi\-hop chains collapse to direct adjacency\.
3. 3\.Decay—aged structure must weaken\. Without it, there is no forgetting; stale connections dominate attention, and “recent” becomes indistinguishable from “active”\. Our operation: all edges undergo exponential decay at every consolidation pass, following MemoryBank’s\(Zhonget al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib15)\)application of the Ebbinghaus curve\.
4. 4\.Completion—connections invisible to a local LLM view must be inferred\. Without it, the LLM sees only the current event plus its context window and cannot know that a concept created now should connect to one created three sessions ago; cross\-session structure fragments into orphans\. Our operation: kNN inference over concept embeddings \(parameters in Appendix[B](https://arxiv.org/html/2605.13438#A2)\) scans for zero\-edge concept nodes and createsGROUNDSconnections—automatically repairing gaps the LLM’s local\-view planning misses\.

Prior systems address each debt at most partially \(Appendix[C](https://arxiv.org/html/2605.13438#A3), Table[12](https://arxiv.org/html/2605.13438#A3.T12)\): HippoRAG covers a narrow form of completion via synonym edges; Mem0 performs write\-time dedup without post\-hoc consolidation or decay; MAGMA’s slow\-path inference densifies an ingested batch but does not reinforce, compress, or decay; A\-Mem and PREMem operate at the text\-rewrite layer and never modify graph structure\.CogniFoldis the first agent\-memory system to address all four debts as automatic, topology\-level operations—the mutually\-reinforcing cycleREINFORCES→\\toMERGE\_NODES→\\tokNN completion, balanced by edge decay\. Why graph\-level and not text\-level or gradient\-level? Text rewriting updates the content of a note while leaving graph\-theoretic distance, PageRank, and reasoning paths invariant; gradient\-based memory updates weights that cannot be inspected, audited, or selectively deleted\(McCloskey and Cohen,[1989](https://arxiv.org/html/2605.13438#bib.bib44)\)\. Only topology change makes memory’s internal geometry both mutable and inspectable\.

### 3\.3Intent Emergence

Intentnodes emerge when concept\-cluster density crosses a threshold\(Einstein and McDaniel,[2005](https://arxiv.org/html/2605.13438#bib.bib4); Gilboa and Marlatte,[2017](https://arxiv.org/html/2605.13438#bib.bib6)\): converging evidence across multiple concepts signals an unmet goal, and the LLM crystallizes it as an intent linked to its supporting concepts viaTRIGGERSedges\. Each intent follows a lifecycle \(pending→\\toresolved\|\|rejected\|\|deferred\) that provides goal\-directed organization for always\-on agents\.

A per\-category EMA loop calibrates the emission threshold from accept/reject/defer/modify feedback:

wc\(t\)=\(1−αema\)⋅wc\(t−1\)\+αema⋅st,w\_\{c\}^\{\(t\)\}=\(1\-\\alpha\_\{\\text\{ema\}\}\)\\cdot w\_\{c\}^\{\(t\-1\)\}\+\\alpha\_\{\\text\{ema\}\}\\cdot s\_\{t\},\(2\)wherests\_\{t\}maps each feedback type to a numeric score \(Appendix[B](https://arxiv.org/html/2605.13438#A2)\)\. Categories the user consistently accepts see lowered thresholds; rejected categories are suppressed—prediction\-error correction\(Friston,[2010](https://arxiv.org/html/2605.13438#bib.bib2); Clark,[2013](https://arxiv.org/html/2605.13438#bib.bib3)\)applied to intent generation rather than model weights\.

Under the single\-session QA protocols of §[4](https://arxiv.org/html/2605.13438#S4), concept\-cluster density never reaches the emission threshold, so intents are not triggered there\. In the controlled multi\-domain streams of CogEval\-Bench \(§[4\.5](https://arxiv.org/html/2605.13438#S4.SS5)\), the threshold is reached repeatedly and intent emission is measured directly \(Proactivity 0\.614\)\.

## 4Experiments and Results

### 4\.1Datasets

Proactive Evaluation\.QA accuracy alone cannot validate the central claim that cognitive structures emerge from event\-stream folding—a flat RAG system with strong BM25 can score well on factual QA without forming any concepts, and a verbatim event store can pass multi\-hop retrieval without any compression\. We therefore introduceCogEval\-Bench, a structural diagnostic benchmark\. CogEval\-Bench usestop\-down generation: for each scenario a gold concept graph𝒢∗=\(𝒞∗,ℛ∗,ℋ∗,ℐ∗\)\\mathcal\{G\}^\{\*\}=\(\\mathcal\{C\}^\{\*\},\\mathcal\{R\}^\{\*\},\\mathcal\{H\}^\{\*\},\\mathcal\{I\}^\{\*\}\)is specified first \(concepts, inter\-concept relations, hierarchy parents, expected intents, and planted multi\-hop chains\), then grounded first\-person events are generated from it, followed by distractor injection \(10–15%\) and temporal shuffling\. The benchmark spans 6 scenarios across 4 domains \(SoftEng, Health, Team, News, Academic, Support\); scale statistics are reported in Appendix[F](https://arxiv.org/html/2605.13438#A6)\. Ground truth is establishedby constructionrather than through post\-hoc annotation\. Three evaluation tracks are computed per system:Concept Emergence\(Gold F1 via Hungarian\-matched\(Kuhn,[1955](https://arxiv.org/html/2605.13438#bib.bib57)\)soft\-matching, LLM Quality, Harmony, Purity\),Relationship Topology\(Chain Discovery, Clustering, Modularity, Edge Type Entropy\), andCompression & Proactivity\(Compression Ratio, PageRank Gini, Proactivity\)\. Full schemas, generation prompts, and per\-scenario breakdowns are in Appendix[F](https://arxiv.org/html/2605.13438#A6)\.

Memory\-Quality Evaluation\.Downstream memory utility is evaluated across 7 broad\-coverage benchmarks spanning 5 cognitive domains: dialogue coherence \(MuTual\(Cuiet al\.,[2020](https://arxiv.org/html/2605.13438#bib.bib21)\)\), theory of mind \(ToMi\(Leet al\.,[2019](https://arxiv.org/html/2605.13438#bib.bib97)\)\), multi\-hop reasoning \(MuSiQue\(Trivediet al\.,[2022](https://arxiv.org/html/2605.13438#bib.bib22)\)\), narrative comprehension \(NarrativeQA\(Kočiskỳet al\.,[2018](https://arxiv.org/html/2605.13438#bib.bib32)\)\), streaming temporal QA \(StreamingQA\(Liskaet al\.,[2022](https://arxiv.org/html/2605.13438#bib.bib26)\)\), conversational memory \(LoCoMo\(Maharanaet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib63)\), full 10\-conversation Mem0 protocol\), and long\-context factual extraction \(BABILong\(Kuratovet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib20)\)\)\. Per\-benchmark sample sizes, baselines, and detailed results are in §[4\.5](https://arxiv.org/html/2605.13438#S4.SS5)\(Table[5](https://arxiv.org/html/2605.13438#S4.T5)and Figure[4](https://arxiv.org/html/2605.13438#S4.F4)\)\.

### 4\.2Baselines

Memory\-Quality baselines\.On LoCoMo, we compare against MIRIX\(Wang and Chen,[2025](https://arxiv.org/html/2605.13438#bib.bib69)\), Mem0\(Chhikaraet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib43)\), Zep\(Rasmussenet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib68)\), Memobase\(MemoDB Team,[2026](https://arxiv.org/html/2605.13438#bib.bib67)\), Supermemory\(Supermemory Team,[2026](https://arxiv.org/html/2605.13438#bib.bib70)\), MemU\(NevaMind AI,[2025](https://arxiv.org/html/2605.13438#bib.bib71)\), MemOS\(Liet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib64)\), and ENGRAM\(Patel and Patel,[2025](https://arxiv.org/html/2605.13438#bib.bib65)\)under the matched single\-judgegpt\-4o\-miniMem0 protocol; numbers come fromLiet al\.\([2025](https://arxiv.org/html/2605.13438#bib.bib64)\)’s public reproduction for all but ENGRAM \(taken fromPatel and Patel,[2025](https://arxiv.org/html/2605.13438#bib.bib65)\) and Zep \(corrected reproduction\(Rasmussenet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib68)\)\)\. On MuSiQue, we adopt the standard graph\-retrieval suite ofGutiérrezet al\.\([2025](https://arxiv.org/html/2605.13438#bib.bib34)\)—BM25, Contriever, NV\-Embed\-v2, RAPTOR\(Sarthiet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib72)\), GraphRAG\(Edgeet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib40)\), LightRAG\(Guoet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib73)\), HippoRAG\(Gutiérrezet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib33)\), HippoRAG 2\(Gutiérrezet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib34)\)—plus PolicyRAG\(Sarnaiket al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib66)\)\. On the remaining benchmarks we report against the most\-cited published baselines under each benchmark’s headline metric \(Figure[4](https://arxiv.org/html/2605.13438#S4.F4)\)\.

Proactive baselines\.On CogEval\-Bench, seven systems are compared under identical LLM and events:OpenIE KG\(HippoRAG\-style triples\),Cognee\(Topoteretes,[2026](https://arxiv.org/html/2605.13438#bib.bib41)\)\(Extract–Cognify–Load pipeline\),HippoRAG 2\(Gutiérrezet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib34)\)\(deep passage \+ synonym expansion\),GraphRAG\(Edgeet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib40)\)\(batch community detection \+ LLM summarisation\),Mem0\(Chhikaraet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib43)\)\(text\-rewrite memory cells\),Zep\(Rasmussenet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib68)\)\(temporal knowledge graph\), andCogniFold\.

### 4\.3Implementation Details

Table 2:CogEval\-Bench: structural evaluation across 7 systems\.Averages over 6 scenarios \(small scale,∼42\{\\sim\}42events each\)\. Track A measures concept quality, Track B measures graph topology, Track C measures compression and proactivity\. Arrows indicate preferred direction\. All systems share GPT\-4o\-mini andtext\-embedding\-3\-small; differences are attributable to architecture\. OnlyCogniFoldachieves non\-zero purity and proactivity—structural properties absent from entity\-level, batch\-processed, or text\-rewrite representations\. Bold: best; underline: second\-best\.†Mem0 has no native graph; we materialise each extracted memory as a node and induce edges via vector\-similarity neighbours, which trivially saturates Chain Disc\. but produces no typed\-edge structure \(Edge Entropy=0=0\)\.

We usegpt\-4o\-minias the agent and reader on every benchmark, withtext\-embedding\-3\-smallas the embedding model throughout, so cross\-system performance differences are attributable to architectural design rather than reader capability\. All benchmarks run with stream ingestion: each event is processed sequentially through the full write path \(context assembly→\\toUpdatePlan→\\toatomic execution→\\toconsolidation\), so the topology\-evolution operations of §[3\.2](https://arxiv.org/html/2605.13438#S3.SS2)fire per event and consolidation operates throughout ingestion—the same online operationCogniFoldperforms in deployment, not a one\-shot batch pass at the end\. Hyper\-parameters, full prompts, and the cost / reproducibility statement are in Appendix[B](https://arxiv.org/html/2605.13438#A2)and Appendix[G](https://arxiv.org/html/2605.13438#A7)\.

### 4\.4Graph Evolution

Table[3](https://arxiv.org/html/2605.13438#S4.T3)reports per\-benchmark statistics of the substrateCogniFoldproduces under the protocol above—events ingested, concepts crystallised, edges accreted, and the compression ratio achieved by the consolidation operations of §[3\.2](https://arxiv.org/html/2605.13438#S3.SS2)\. The numbers serve a diagnostic role: they make explicit the structural footprint each downstream score in §[4\.5](https://arxiv.org/html/2605.13438#S4.SS5)is sitting on, so the reader can read accuracy alongside the graph that produced it\.

Table 3:Graph evolution statistics per benchmark\.Compression = concepts/events \(lower==more folding\)\. Edge density = edges/\(concepts\+intents\), measuring connectivity of higher\-level nodes\. Statistics are averaged over all ingested samples per benchmark\. Per\-benchmark accuracy results are in Table[4](https://arxiv.org/html/2605.13438#S4.T4)\(LoCoMo\) and Figure[4](https://arxiv.org/html/2605.13438#S4.F4)\(the other six\)\.BenchmarkEventsConceptsIntentsEdgesCompressionDensityMuTual42182350\.431\.75ToMi1285451120\.421\.88BABILong48019583100\.411\.53StreamingQA620340124850\.551\.38MuSiQue38026063200\.681\.20LoCoMo850520186800\.611\.22NarrativeQA\(Kočiskỳet al\.,[2018](https://arxiv.org/html/2605.13438#bib.bib32)\)1,200480221,6500\.403\.29
### 4\.5Results

Proactive Results\.

Table[2](https://arxiv.org/html/2605.13438#S4.T2)reports CogEval\-Bench averages across all six scenarios; per\-scenario breakdowns and full evaluation details are in Appendix[F](https://arxiv.org/html/2605.13438#A6)\.CogniFoldachieves Harmony 0\.476, substantially above GraphRAG \(0\.323, the strongest baseline\) and far above entity\-level systems \(OpenIE KG 0\.138; Cognee 0\.094; HippoRAG 2 0\.095\)\.CogniFoldis the only system producing non\-zero Purity \(0\.361\)—its concepts are coherently grounded in their constituent events, while all baselines lack event\-level grounding\. On topology,CogniFold’s clustering coefficient \(0\.327\) reflects genuine triadic closure among semantically related concepts, distinct from HippoRAG 2’s high raw clustering \(0\.716\) that arises from synonym\-expansion cliques among name variants\. On compression and proactivity,CogniFoldachieves4\.6×4\.6\\timescompression \(41–47 events→\\to7–12 concepts\) while OpenIE KG and HippoRAG 2 expand the representation; andCogniFoldis the only system emitting intent nodes at all, reaching Proactivity 0\.614 \(61% of intents grounded by≥2\\geq 2supporting connections\)\.

The seven\-system comparison reveals an ordered hierarchy in representational richness:entity graphs\(OpenIE KG\)—fragmented triples, high modularity from disconnection;entity graphs with enrichment\(Cognee, HippoRAG 2\)—shallow structure atop entity extraction, Harmony stuck at∼0\.09\{\\sim\}0\.09;community graphs\(GraphRAG\)—batch community detection yields the strongest baseline \(Harmony 0\.323\) but with zero event\-level grounding \(Purity=0=0\) and negligible clustering \(0\.002\);cognitive graphs\(CogniFold\)—online folding with merging yields event\-grounded concepts, genuine triadic closure, substantial compression, and proactive goal identification simultaneously\. The critical architectural distinction isonline, incremental processing with merging: Cognee and HippoRAG 2 add machinery \(ECL, synonym expansion, PPR\) without cross\-event integration and remain at the entity\-enrichment tier, confirming that the bottleneck in concept emergence is not extraction depth but the ability to recognise that eventse1,e3,e7e\_\{1\},e\_\{3\},e\_\{7\}ground the same underlying concept and merge them into a single abstraction\. This hierarchy parallels a neuroscience progression\(McClellandet al\.,[1995](https://arxiv.org/html/2605.13438#bib.bib1); Gilboa and Marlatte,[2017](https://arxiv.org/html/2605.13438#bib.bib6); Preston and Eichenbaum,[2013](https://arxiv.org/html/2605.13438#bib.bib61)\): episodic storage, pattern separation without consolidation, shallow categorisation, schema extraction, and active consolidation with goal generation\.

Memory\-Quality Results\.

Table 4:LoCoMo per\-category and aggregate comparison\.J\-Score is the LLM\-as\-judge accuracy\(Zhenget al\.,[2023](https://arxiv.org/html/2605.13438#bib.bib77)\)withgpt\-4o\-minias the judge; Tokens is per\-question input\+output token consumption\.Bold: best per column;underline: second\-best\.†Reported under a 3\-LLM\-judge ensemble protocol; numbers as reported inHuet al\.\([2026](https://arxiv.org/html/2605.13438#bib.bib96)\)Table 1\.

CogEval\-Bench tells us that the substrate forms cognitive structure; the seven downstream benchmarks of our suite \(Table[5](https://arxiv.org/html/2605.13438#S4.T5)\) ask whether that structure pays off as memory\. The sameCogniFoldgraph—built once per task by the same write path, queried by the same read path, evaluated under the samegpt\-4o\-minireader—is set against eight memory systems on the conversational\-memory benchmark \(Table[4](https://arxiv.org/html/2605.13438#S4.T4), full 10\-conversation Mem0 protocol with matched judge\), against the standard graph\-retrieval suite ofGutiérrezet al\.\([2025](https://arxiv.org/html/2605.13438#bib.bib34)\)on the multi\-hop reasoning benchmark, and against the most\-cited published baselines on each of the remaining five; Figure[4](https://arxiv.org/html/2605.13438#S4.F4)summarises the six per\-benchmark comparisons at a glance\. On LoCoMo the substrate leads the audit\-resilient region of the leaderboard\(Penfield Labs,[2026](https://arxiv.org/html/2605.13438#bib.bib78)\), scoring above MemOS, ENGRAM, and the text\-rewriting tier; on MuSiQue it reaches F1 58\.7, exceeding the strongest published RAG pipeline \(HippoRAG 2, \+9\.4\) and the strongest symbolic\-graph alternative \(PolicyRAG, \+2\.8\); on the remaining five it leads on theory of mind \(ToMi, \+3\.3 over AutoToM\) and long\-context factual extraction \(BABILong, \+1\.2 over fine\-tuned ARMT\), holds within range of streaming\-FiD on StreamingQA, and tops the published memory and structure\-augmented baselines on MuTual and NarrativeQA\.

Table 5:Downstream benchmark suite\.Seven benchmarks across five cognitive domains\. Per\-benchmark detailed system comparisons are reported in Table[4](https://arxiv.org/html/2605.13438#S4.T4)\(LoCoMo\) and Figure[4](https://arxiv.org/html/2605.13438#S4.F4)\(the other six\)\.![Refer to caption](https://arxiv.org/html/2605.13438v1/x5.png)Figure 4:Downstream benchmarks at a glance\.CogniFold \(indigo, bold\) against the most\-cited published baselines for each benchmark, sorted by score with the best on top\. Metric varies per benchmark; sample sizes are 500 for MuSiQue, NarrativeQA, MuTual, StreamingQA, ToMi, and 100 for BABILong\.MuSiQue \(multi\-hop reasoning\)\.On MuSiQueCogniFoldis benchmarked against the standard graph\-augmented retrieval suite ofGutiérrezet al\.\([2025](https://arxiv.org/html/2605.13438#bib.bib34)\)—RAPTOR, GraphRAG, HippoRAG—together with PolicyRAG\(Sarnaiket al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib66)\), all under a unifiedgpt\-4o\-minireader, and reaches F1 58\.7 / EM 48\.0, the highest of the suite under this reader\. The lift over HippoRAG \(\+9\.4 F1\) and PolicyRAG \(\+2\.8 F1\) reflects the formation\-over\-retrieval principle: a dynamically folded graph carries more answer\-relevant structure than a static graph paired with a sophisticated retrieval pipeline\.

Per\-benchmark summary on the remaining five\.MuTual\(dialogue coherence\) gains arise from concept\-level folding of recurring conversational themes that paragraph\-level chunking destroys;ToMi\(theory of mind\) gains come from the symbolic belief tracker \(§LABEL:sec:query\) bypassing LLM over\-abstraction of spatial state;NarrativeQA\(long\-form fiction\) benefits from cross\-character entity disambiguation enabled by concept folding;StreamingQA\(time\-anchored facts\) benefits from explicit time nodes that allow temporal queries to resolve by traversal rather than vector ranking;BABILong\(long\-context bAbI\-style\) is the one regime where structural memory is approximately neutral—local key–value supports are retrievable directly without need of cross\-event integration\.

What the same substrate doing well across these very different cognitive domains shows is not a tuning result but a generality result\. Conversational memory, multi\-hop reasoning, theory of mind, narrative comprehension, and streaming temporal QA pull on different cognitive operations—inter\-session consolidation, cross\-document chaining, belief tracking, character disambiguation, time\-anchored recall—yet they share the same underlying ask: that the memory substrate retain the right relational structure between events and surface it on demand\.CogniFold’s consistent placement in the upper band across this spread is direct evidence that*cognitive folding*—the operations of §[3\.2](https://arxiv.org/html/2605.13438#S3.SS2)—is a task\-general write\-path competence rather than a benchmark\-tuned heuristic\. The one regime where the gain disappears is exactly the one the theory predicts: BABILong asks for verbatim local key–value supports, where structural folding adds no leverage that direct retrieval did not already have\.

## 5Discussion

In a strict sense,CogniFoldis beyond memory: it does not merely store and retrieve, it bootstraps\. Each fold collapses events into higher\-level concepts that later folds reason against; the agent is not starting cold at every event but using its own accumulated cognition as substrate\. The 4\.6×\\timescompression and 0\.614 Proactivity \(§[4\.5](https://arxiv.org/html/2605.13438#S4.SS5)\) are the architectural correlates of this process, reflecting both the aggressive folding of events and how bootstrapped concepts further crystallize into goals\.

However, this bootstrapping dynamic introduces a principled limitation:path\-dependence\. Because each fold conditions on previously accumulated structure, the same events in different orders produce different graphs\. While this mirrors humancurriculum effects—where pedagogical ordering yields cleaner schemas than shuffled inputs\(Elman,[1993](https://arxiv.org/html/2605.13438#bib.bib62); Zhaoet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib83)\)—it raises an open question regarding memory stability\. Order\-aware consolidation, replay\-based smoothing, and bounded\-divergence analyses on streaming graphs are concrete avenues for future work\.

A second limitation concerns the depth of our prefrontal mapping\.CogniFoldoperationalizes only schema\-driven integration, whereas the biological prefrontal cortex performs reward\-based valuation, cognitive control, and counterfactual simulation\(Gilboa and Marlatte,[2017](https://arxiv.org/html/2605.13438#bib.bib6); Preston and Eichenbaum,[2013](https://arxiv.org/html/2605.13438#bib.bib61)\)\. Without value estimation, it cannot rank intents by long\-horizon utility; without cognitive control, it cannot suppress impulsive emissions when a stronger goal is active; without counterfactual rollout, it cannot anticipate downstream consequences\. Integrating these mechanisms forms the natural research arc beyond this paper\.

## 6Conclusion

We presentCogniFold, an always\-on proactive agent memory that folds fragmented events into persistent cognitive structure\. Unlike reactive retrieval systems,CogniFoldbuilds a living graph that continuously folds, merges, decays, and reconnects under the event stream\. Because cognition grows recursively from the system’s own products, goal\-directed intents naturally emerge from converging evidence\. We validated this design across two critical axes: proactive structural emergence on CogEval\-Bench, and robust memory quality across seven broad\-coverage downstream benchmarks\.

As foundation\-model capability grows, what a system computes in a single forward pass approaches what a human can reason about in a moment; what it accumulates, organizes, and bootstraps across time is where the value of an always\-on agent will increasingly accrue\. We releaseCogniFoldas a foundation for research on real\-time interaction, proactive collaboration, and agent cognition that bootstraps beyond memory\.

## References

- Remembering: a study in experimental and social psychology\.Cambridge university press\.Cited by:[§F\.1](https://arxiv.org/html/2605.13438#A6.SS1.p1.8),[Table 1](https://arxiv.org/html/2605.13438#S2.T1.1.8.6.3),[item 1](https://arxiv.org/html/2605.13438#S3.I1.i1.p1.1)\.
- A\. Behrouz, P\. Zhong, and V\. Mirrokni \(2024\)Titans: learning to memorize at test time\.arXiv preprint arXiv:2501\.00663\.Cited by:[§A\.2](https://arxiv.org/html/2605.13438#A1.SS2.p1.1),[Table 6](https://arxiv.org/html/2605.13438#A1.T6.14.7.6.1),[§1](https://arxiv.org/html/2605.13438#S1.p3.1)\.
- B\. Bontempi, C\. Laurent\-Demir, C\. Destrade, and R\. Jaffard \(1999\)Time\-dependent reorganization of brain circuitry underlying long\-term memory storage\.Nature400\(6745\),pp\. 671–675\.Cited by:[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p1.1)\.
- M\. Bratman \(1987\)Intention, plans, and practical reason\.Harvard University Press\.Cited by:[§A\.3](https://arxiv.org/html/2605.13438#A1.SS3.p1.1),[§1](https://arxiv.org/html/2605.13438#S1.p3.1)\.
- S\. Carey \(2000\)The origin of concepts\.Journal of Cognition and Development1\(1\),pp\. 37–41\.Cited by:[§1](https://arxiv.org/html/2605.13438#S1.p3.1),[§2\.2](https://arxiv.org/html/2605.13438#S2.SS2.p1.1)\.
- P\. Chhikara, D\. Khant, S\. Aryan, T\. Singh, and D\. Yadav \(2025\)Mem0: building production\-ready ai agents with scalable long\-term memory\.arXiv preprint arXiv:2504\.19413\.Cited by:[§A\.1](https://arxiv.org/html/2605.13438#A1.SS1.p1.1),[Table 6](https://arxiv.org/html/2605.13438#A1.T6.14.3.2.1),[Table 12](https://arxiv.org/html/2605.13438#A3.T12.1.5.4.1),[§F\.5](https://arxiv.org/html/2605.13438#A6.SS5.p1.2),[§1](https://arxiv.org/html/2605.13438#S1.p1.1),[§1](https://arxiv.org/html/2605.13438#S1.p2.1),[§3\.1](https://arxiv.org/html/2605.13438#S3.SS1.p5.1),[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1),[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p2.1),[Table 4](https://arxiv.org/html/2605.13438#S4.T4.1.4.2.1)\.
- A\. Clark \(2013\)Whatever next? predictive brains, situated agents, and the future of cognitive science\.Behavioral and brain sciences36\(3\),pp\. 181–204\.Cited by:[§3\.3](https://arxiv.org/html/2605.13438#S3.SS3.p2.1)\.
- L\. Cui, Y\. Wu, S\. Liu, Y\. Zhang, and M\. Zhou \(2020\)MuTual: a dataset for multi\-turn dialogue reasoning\.InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics,pp\. 1406–1416\.Cited by:[Appendix G](https://arxiv.org/html/2605.13438#A7.p1.1),[§4\.1](https://arxiv.org/html/2605.13438#S4.SS1.p2.1),[Table 5](https://arxiv.org/html/2605.13438#S4.T5.1.6.5.1)\.
- A\. F\. de Sousa, Z\. E\. Zeidler, D\. G\. Almeida\-Filho, Y\. Shen, A\. Luchetti, S\. Simanian, M\. Mardini, L\. A\. DeNardo, and A\. J\. Silva \(2026\)The prefrontal cortex controls memory organization in the hippocampus\.Nature Neuroscience,pp\. 1–12\.Cited by:[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p2.1)\.
- H\. Ebbinghaus \(2013\)\[Image\] memory: a contribution to experimental psychology\.Annals of neurosciences20\(4\),pp\. 155\.Cited by:[§3\.1](https://arxiv.org/html/2605.13438#S3.SS1.p4.1)\.
- D\. Edge, H\. Trinh, N\. Cheng, J\. Bradley, A\. Chao, A\. Mody, S\. Truitt, D\. Metropolitansky, R\. O\. Ness, and J\. Larson \(2024\)From local to global: a graph rag approach to query\-focused summarization\.arXiv preprint arXiv:2404\.16130\.Cited by:[§F\.5](https://arxiv.org/html/2605.13438#A6.SS5.p1.2),[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1),[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p2.1)\.
- H\. Eichenbaum \(2017\)Memory: organization and control\.Annual review of psychology68,pp\. 19–45\.Cited by:[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p2.1),[Table 1](https://arxiv.org/html/2605.13438#S2.T1.1.7.5.3)\.
- G\. O\. Einstein and M\. A\. McDaniel \(2005\)Prospective memory: multiple retrieval processes\.Current Directions in Psychological Science14\(6\),pp\. 286–290\.Cited by:[§A\.3](https://arxiv.org/html/2605.13438#A1.SS3.p1.1),[§1](https://arxiv.org/html/2605.13438#S1.p1.1),[Table 1](https://arxiv.org/html/2605.13438#S2.T1.1.5.3.3),[Table 1](https://arxiv.org/html/2605.13438#S2.T1.1.9.7.3),[§3\.3](https://arxiv.org/html/2605.13438#S3.SS3.p1.3)\.
- J\. L\. Elman \(1993\)Learning and development in neural networks: the importance of starting small\.Cognition48\(1\),pp\. 71–99\.Cited by:[§5](https://arxiv.org/html/2605.13438#S5.p2.1)\.
- P\. W\. Frankland and B\. Bontempi \(2005\)The organization of recent and remote memories\.Nature reviews neuroscience6\(2\),pp\. 119–130\.Cited by:[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p1.1)\.
- K\. Friston \(2010\)The free\-energy principle: a unified brain theory?\.Nature reviews neuroscience11\(2\),pp\. 127–138\.Cited by:[§3\.3](https://arxiv.org/html/2605.13438#S3.SS3.p2.1)\.
- A\. Gilboa and H\. Marlatte \(2017\)Neurobiology of schemas and schema\-mediated memory\.Trends in cognitive sciences21\(8\),pp\. 618–631\.Cited by:[§F\.1](https://arxiv.org/html/2605.13438#A6.SS1.p1.8),[item 2](https://arxiv.org/html/2605.13438#S3.I1.i2.p1.1),[§3\.3](https://arxiv.org/html/2605.13438#S3.SS3.p1.3),[§4\.5](https://arxiv.org/html/2605.13438#S4.SS5.p3.3),[§5](https://arxiv.org/html/2605.13438#S5.p3.1)\.
- Z\. Guo, L\. Xia, Y\. Yu, T\. Ao, and C\. Huang \(2024\)Lightrag: simple and fast retrieval\-augmented generation\.arXiv preprint arXiv:2410\.057792\(3\)\.Cited by:[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1)\.
- B\. J\. Gutiérrez, Y\. Shu, Y\. Gu, M\. Yasunaga, and Y\. Su \(2024\)Hipporag: neurobiologically inspired long\-term memory for large language models\.Advances in neural information processing systems37,pp\. 59532–59569\.Cited by:[§A\.1](https://arxiv.org/html/2605.13438#A1.SS1.p1.1),[Table 6](https://arxiv.org/html/2605.13438#A1.T6.14.2.1.1),[Table 12](https://arxiv.org/html/2605.13438#A3.T12.1.3.2.1),[§1](https://arxiv.org/html/2605.13438#S1.p2.1),[§3\.1](https://arxiv.org/html/2605.13438#S3.SS1.p3.1),[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1)\.
- B\. J\. Gutiérrez, Y\. Shu, W\. Qi, S\. Zhou, and Y\. Su \(2025\)From rag to memory: non\-parametric continual learning for large language models\.arXiv preprint arXiv:2502\.14802\.Cited by:[§A\.1](https://arxiv.org/html/2605.13438#A1.SS1.p1.1),[Table 6](https://arxiv.org/html/2605.13438#A1.T6.14.2.1.1),[Table 12](https://arxiv.org/html/2605.13438#A3.T12.1.3.2.1),[§F\.5](https://arxiv.org/html/2605.13438#A6.SS5.p1.2),[§1](https://arxiv.org/html/2605.13438#S1.p2.1),[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1),[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p2.1),[§4\.5](https://arxiv.org/html/2605.13438#S4.SS5.p5.1),[§4\.5](https://arxiv.org/html/2605.13438#S4.SS5.p6.1)\.
- D\. O\. Hebb \(2005\)The organization of behavior: a neuropsychological theory\.Psychology press\.Cited by:[Table 1](https://arxiv.org/html/2605.13438#S2.T1.1.6.4.3),[§3\.1](https://arxiv.org/html/2605.13438#S3.SS1.p5.1)\.
- C\. Hu, X\. Gao, Z\. Zhou, D\. Xu, Y\. Bai, X\. Li, H\. Zhang, T\. Li, C\. Zhang, L\. Bing, and Y\. Deng \(2026\)EverMemOS: a self\-organizing memory operating system for structured long\-horizon reasoning\.arXiv preprint arXiv:2601\.02163\.Cited by:[Table 4](https://arxiv.org/html/2605.13438#S4.T4.1.1.1),[Table 4](https://arxiv.org/html/2605.13438#S4.T4.2.2)\.
- D\. Jiang, Y\. Li, G\. Li, and B\. Li \(2026\)MAGMA: a multi\-graph based agentic memory architecture for ai agents\.arXiv preprint arXiv:2601\.03236\.Cited by:[§A\.1](https://arxiv.org/html/2605.13438#A1.SS1.p1.1),[Table 6](https://arxiv.org/html/2605.13438#A1.T6.14.4.3.1),[Table 12](https://arxiv.org/html/2605.13438#A3.T12.1.6.5.1),[§1](https://arxiv.org/html/2605.13438#S1.p2.1)\.
- T\. Kočiskỳ, J\. Schwarz, P\. Blunsom, C\. Dyer, K\. M\. Hermann, G\. Melis, and E\. Grefenstette \(2018\)The narrativeqa reading comprehension challenge\.Transactions of the Association for Computational Linguistics6,pp\. 317–328\.Cited by:[Appendix G](https://arxiv.org/html/2605.13438#A7.p1.1),[§4\.1](https://arxiv.org/html/2605.13438#S4.SS1.p2.1),[Table 3](https://arxiv.org/html/2605.13438#S4.T3.6.8.8.1),[Table 5](https://arxiv.org/html/2605.13438#S4.T5.1.4.3.1)\.
- H\. W\. Kuhn \(1955\)The hungarian method for the assignment problem\.Naval research logistics quarterly2\(1\-2\),pp\. 83–97\.Cited by:[§F\.4](https://arxiv.org/html/2605.13438#A6.SS4.p1.2),[§4\.1](https://arxiv.org/html/2605.13438#S4.SS1.p1.1)\.
- D\. Kumaran, D\. Hassabis, and J\. L\. McClelland \(2016\)What learning systems do intelligent agents need? complementary learning systems theory updated\.Trends in cognitive sciences20\(7\),pp\. 512–534\.Cited by:[§1](https://arxiv.org/html/2605.13438#S1.p3.1),[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p1.1)\.
- Y\. Kuratov, A\. Bulatov, P\. Anokhin, I\. Rodkin, D\. Sorokin, A\. Sorokin, and M\. Burtsev \(2024\)Babilong: testing the limits of llms with long context reasoning\-in\-a\-haystack\.Advances in Neural Information Processing Systems37,pp\. 106519–106554\.Cited by:[Appendix G](https://arxiv.org/html/2605.13438#A7.p1.1),[§4\.1](https://arxiv.org/html/2605.13438#S4.SS1.p2.1),[Table 5](https://arxiv.org/html/2605.13438#S4.T5.1.8.7.1)\.
- C\. A\. Kurby and J\. M\. Zacks \(2008\)Segmentation in the perception and memory of events\.Trends in cognitive sciences12\(2\),pp\. 72–79\.Cited by:[§1](https://arxiv.org/html/2605.13438#S1.p1.1)\.
- M\. Le, Y\. Boureau, and M\. Nickel \(2019\)Revisiting the evaluation of theory of mind through question answering\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\),Hong Kong, China,pp\. 5872–5877\.External Links:[Link](https://www.aclweb.org/anthology/D19-1598),[Document](https://dx.doi.org/10.18653/v1/D19-1598)Cited by:[Appendix G](https://arxiv.org/html/2605.13438#A7.p1.1),[§4\.1](https://arxiv.org/html/2605.13438#S4.SS1.p2.1),[Table 5](https://arxiv.org/html/2605.13438#S4.T5.1.7.6.1)\.
- Z\. Li, C\. Xi, C\. Li, D\. Chen, B\. Chen, S\. Song, S\. Niu, H\. Wang, J\. Yang, C\. Tang,et al\.\(2025\)Memos: a memory os for ai system\.arXiv preprint arXiv:2507\.03724\.Cited by:[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1),[Table 4](https://arxiv.org/html/2605.13438#S4.T4.1.9.7.1)\.
- A\. Liska, T\. Kocisky, E\. Gribovskaya, T\. Terzi, E\. Sezener, D\. Agrawal, C\. D\. M\. D’Autume, T\. Scholtes, M\. Zaheer, S\. Young,et al\.\(2022\)Streamingqa: a benchmark for adaptation to new knowledge over time in question answering models\.InInternational Conference on Machine Learning,pp\. 13604–13622\.Cited by:[Appendix G](https://arxiv.org/html/2605.13438#A7.p1.1),[§4\.1](https://arxiv.org/html/2605.13438#S4.SS1.p2.1),[Table 5](https://arxiv.org/html/2605.13438#S4.T5.1.5.4.1)\.
- A\. Maharana, D\. Lee, S\. Tulyakov, M\. Bansal, F\. Barbieri, and Y\. Fang \(2024\)Evaluating very long\-term conversational memory of llm agents\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 13851–13870\.Cited by:[Appendix G](https://arxiv.org/html/2605.13438#A7.p1.1),[§4\.1](https://arxiv.org/html/2605.13438#S4.SS1.p2.1),[Table 5](https://arxiv.org/html/2605.13438#S4.T5.1.2.1.1)\.
- B\. P\. Majumder, B\. D\. Mishra, P\. Jansen, O\. Tafjord, N\. Tandon, L\. Zhang, C\. Callison\-Burch, and P\. Clark \(2023\)Clin: a continually learning language agent for rapid task adaptation and generalization\.arXiv preprint arXiv:2310\.10134\.Cited by:[§1](https://arxiv.org/html/2605.13438#S1.p1.1)\.
- D\. Marr \(1971\)Simple memory: a theory for archicortex\.Philosophical Transactions of the Royal Society of London\. B, Biological Sciences262\(841\),pp\. 23–81\.Cited by:[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p1.1)\.
- J\. L\. McClelland, B\. L\. McNaughton, and R\. C\. O’Reilly \(1995\)Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory\.\.Psychological review102\(3\),pp\. 419\.Cited by:[§1](https://arxiv.org/html/2605.13438#S1.p3.1),[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p1.1),[Table 1](https://arxiv.org/html/2605.13438#S2.T1.1.10.8.3),[Table 1](https://arxiv.org/html/2605.13438#S2.T1.1.3.1.3),[§4\.5](https://arxiv.org/html/2605.13438#S4.SS5.p3.3)\.
- M\. McCloskey and N\. J\. Cohen \(1989\)Catastrophic interference in connectionist networks: the sequential learning problem\.InPsychology of learning and motivation,Vol\.24,pp\. 109–165\.Cited by:[§A\.2](https://arxiv.org/html/2605.13438#A1.SS2.p1.1),[§3\.2](https://arxiv.org/html/2605.13438#S3.SS2.p3.2)\.
- MemoDB Team \(2026\)Memobase: user profile\-based long\-term memory for AI chatbot applications\.Note:[https://github\.com/memodb\-io/memobase](https://github.com/memodb-io/memobase)Version 0\.0\.18Cited by:[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1),[Table 4](https://arxiv.org/html/2605.13438#S4.T4.1.6.4.1)\.
- G\. A\. Miller \(1956\)The magical number seven, plus or minus two: some limits on our capacity for processing information\.\.Psychological review63\(2\),pp\. 81\.Cited by:[§3\.1](https://arxiv.org/html/2605.13438#S3.SS1.p1.1)\.
- NevaMind AI \(2025\)MemU: a memory operating system for agents\.Note:[https://github\.com/NevaMind\-AI/memU](https://github.com/NevaMind-AI/memU)Cited by:[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1),[Table 4](https://arxiv.org/html/2605.13438#S4.T4.1.8.6.1)\.
- M\. E\. Newman \(2006\)Modularity and community structure in networks\.Proceedings of the national academy of sciences103\(23\),pp\. 8577–8582\.Cited by:[§F\.4](https://arxiv.org/html/2605.13438#A6.SS4.p1.2)\.
- R\. C\. O’Reilly and J\. L\. McClelland \(1994\)Hippocampal conjunctive encoding, storage, and recall: avoiding a trade\-off\.Hippocampus4\(6\),pp\. 661–682\.Cited by:[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p1.1)\.
- C\. Packer, V\. Fang, S\. Patil, K\. Lin, S\. Wooders, and J\. Gonzalez \(2023\)MemGPT: towards llms as operating systems\.\.arXiv preprint arXiv:2310\.08560\.Cited by:[§A\.2](https://arxiv.org/html/2605.13438#A1.SS2.p1.1),[§1](https://arxiv.org/html/2605.13438#S1.p1.1)\.
- J\. S\. Park, J\. O’Brien, C\. J\. Cai, M\. R\. Morris, P\. Liang, and M\. S\. Bernstein \(2023\)Generative agents: interactive simulacra of human behavior\.InProceedings of the 36th annual acm symposium on user interface software and technology,pp\. 1–22\.Cited by:[§A\.2](https://arxiv.org/html/2605.13438#A1.SS2.p1.1)\.
- D\. Patel and S\. Patel \(2025\)Engram: effective, lightweight memory orchestration for conversational agents\.arXiv preprint arXiv:2511\.12960\.Cited by:[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1),[Table 4](https://arxiv.org/html/2605.13438#S4.T4.1.10.8.1)\.
- K\. Patterson, P\. J\. Nestor, and T\. T\. Rogers \(2007\)Where do you know what you know? the representation of semantic knowledge in the human brain\.Nature reviews neuroscience8\(12\),pp\. 976–987\.Cited by:[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p1.1)\.
- Penfield Labs \(2026\)Auditing LoCoMo: 6\.4% answer\-key error rate, judge leniency, and reproducibility failures in long\-term conversational memory benchmarks\.Note:[https://github\.com/dial481/locomo\-audit](https://github.com/dial481/locomo-audit)Cited by:[§4\.5](https://arxiv.org/html/2605.13438#S4.SS5.p5.1)\.
- A\. R\. Preston and H\. Eichenbaum \(2013\)Interplay of hippocampus and prefrontal cortex in memory\.Current biology23\(17\),pp\. R764–R773\.Cited by:[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p2.1),[§4\.5](https://arxiv.org/html/2605.13438#S4.SS5.p3.3),[§5](https://arxiv.org/html/2605.13438#S5.p3.1)\.
- P\. Rasmussen, P\. Paliychuk, T\. Beauvais, J\. Ryan, and D\. Chalef \(2025\)Zep: a temporal knowledge graph architecture for agent memory\.arXiv preprint arXiv:2501\.13956\.Cited by:[§A\.1](https://arxiv.org/html/2605.13438#A1.SS1.p1.1),[Table 6](https://arxiv.org/html/2605.13438#A1.T6.14.5.4.1),[Table 12](https://arxiv.org/html/2605.13438#A3.T12.1.1.2),[§F\.5](https://arxiv.org/html/2605.13438#A6.SS5.p1.2),[§1](https://arxiv.org/html/2605.13438#S1.p2.1),[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1),[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p2.1),[Table 4](https://arxiv.org/html/2605.13438#S4.T4.1.5.3.1)\.
- T\. Sarnaik, M\. Shah, and R\. Hegde \(2025\)PolicyRAG: prompt\-guided symbolic graph memory for interpretable multi\-hop retrieval\.Note:[https://openreview\.net/forum?id=0xlI09pvBs](https://openreview.net/forum?id=0xlI09pvBs)Cited by:[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1),[§4\.5](https://arxiv.org/html/2605.13438#S4.SS5.p6.1)\.
- P\. Sarthi, S\. Abdullah, A\. Tuli, S\. Khanna, A\. Goldie, and C\. D\. Manning \(2024\)Raptor: recursive abstractive processing for tree\-organized retrieval\.InThe Twelfth International Conference on Learning Representations,Cited by:[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1)\.
- N\. Shinn, F\. Cassano, A\. Gopinath, K\. Narasimhan, and S\. Yao \(2023\)Reflexion: language agents with verbal reinforcement learning\.Advances in neural information processing systems36,pp\. 8634–8652\.Cited by:[§A\.3](https://arxiv.org/html/2605.13438#A1.SS3.p1.1),[§1](https://arxiv.org/html/2605.13438#S1.p1.1),[§1](https://arxiv.org/html/2605.13438#S1.p2.1)\.
- L\. R\. Squire \(1992\)Memory and the hippocampus: a synthesis from findings with rats, monkeys, and humans\.\.Psychological review99\(2\),pp\. 195\.Cited by:[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p1.1)\.
- R\. Stickgold and M\. P\. Walker \(2007\)Sleep\-dependent memory consolidation and reconsolidation\.Sleep medicine8\(4\),pp\. 331–343\.Cited by:[§3\.2](https://arxiv.org/html/2605.13438#S3.SS2.p1.1)\.
- T\. Sumers, S\. Yao, K\. R\. Narasimhan, and T\. L\. Griffiths \(2023\)Cognitive architectures for language agents\.Transactions on Machine Learning Research\.Cited by:[§A\.3](https://arxiv.org/html/2605.13438#A1.SS3.p1.1),[§1](https://arxiv.org/html/2605.13438#S1.p1.1)\.
- Supermemory Team \(2026\)Supermemory: state\-of\-the\-art memory and context engine for ai\.Note:[https://github\.com/supermemoryai/supermemory](https://github.com/supermemoryai/supermemory)Cited by:[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1),[Table 4](https://arxiv.org/html/2605.13438#S4.T4.1.7.5.1)\.
- Topoteretes \(2026\)Cognee: memory control plane for ai agents\.Note:[https://github\.com/topoteretes/cognee](https://github.com/topoteretes/cognee)Cited by:[Table 12](https://arxiv.org/html/2605.13438#A3.T12.1.4.3.1),[§F\.5](https://arxiv.org/html/2605.13438#A6.SS5.p1.2),[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p2.1)\.
- H\. Trivedi, N\. Balasubramanian, T\. Khot, and A\. Sabharwal \(2022\)♫ MuSiQue: multihop questions via single\-hop question composition\.Transactions of the Association for Computational Linguistics10,pp\. 539–554\.Cited by:[§F\.1](https://arxiv.org/html/2605.13438#A6.SS1.p1.8),[Appendix G](https://arxiv.org/html/2605.13438#A7.p1.1),[§4\.1](https://arxiv.org/html/2605.13438#S4.SS1.p2.1),[Table 5](https://arxiv.org/html/2605.13438#S4.T5.1.3.2.1)\.
- D\. Tse, R\. F\. Langston, M\. Kakeyama, I\. Bethus, P\. A\. Spooner, E\. R\. Wood, M\. P\. Witter, and R\. G\. Morris \(2007\)Schemas and memory consolidation\.Science316\(5821\),pp\. 76–82\.Cited by:[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p2.1)\.
- E\. Tulvinget al\.\(1972\)Episodic and semantic memory\.Organization of memory1\(381\-403\),pp\. 1\.Cited by:[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p1.1)\.
- M\. T\. Van Kesteren, D\. J\. Ruiter, G\. Fernández, and R\. N\. Henson \(2012\)How schema and novelty augment memory formation\.Trends in neurosciences35\(4\),pp\. 211–219\.Cited by:[§2\.1](https://arxiv.org/html/2605.13438#S2.SS1.p2.1)\.
- G\. Wang, Y\. Xie, Y\. Jiang, A\. Mandlekar, C\. Xiao, Y\. Zhu, L\. Fan, and A\. Anandkumar \(2023\)Voyager: an open\-ended embodied agent with large language models\.arXiv preprint arXiv:2305\.16291\.Cited by:[§1](https://arxiv.org/html/2605.13438#S1.p2.1)\.
- Y\. Wang and X\. Chen \(2025\)Mirix: multi\-agent memory system for llm\-based agents\.arXiv preprint arXiv:2507\.07957\.Cited by:[§4\.2](https://arxiv.org/html/2605.13438#S4.SS2.p1.1),[Table 4](https://arxiv.org/html/2605.13438#S4.T4.1.3.1.1)\.
- W\. Xu, Z\. Liang, K\. Mei, H\. Gao, J\. Tan, and Y\. Zhang \(2025\)A\-mem: agentic memory for llm agents\.arXiv preprint arXiv:2502\.12110\.Cited by:[§A\.2](https://arxiv.org/html/2605.13438#A1.SS2.p1.1),[Table 6](https://arxiv.org/html/2605.13438#A1.T6.14.6.5.1),[Table 12](https://arxiv.org/html/2605.13438#A3.T12.1.7.6.1),[§1](https://arxiv.org/html/2605.13438#S1.p2.1),[§1](https://arxiv.org/html/2605.13438#S1.p3.1)\.
- J\. Yang, C\. E\. Jimenez, A\. Wettig, K\. Lieret, S\. Yao, K\. Narasimhan, and O\. Press \(2024\)Swe\-agent: agent\-computer interfaces enable automated software engineering\.Advances in Neural Information Processing Systems37,pp\. 50528–50652\.Cited by:[§1](https://arxiv.org/html/2605.13438#S1.p2.1)\.
- J\. M\. Zacks \(2020\)Event perception and memory\.Annual review of psychology71\(1\),pp\. 165–191\.Cited by:[§1](https://arxiv.org/html/2605.13438#S1.p1.1)\.
- B\. Zhao, C\. G\. Lucas, and N\. R\. Bramley \(2024\)A model of conceptual bootstrapping in human cognition\.Nature Human Behaviour8\(1\),pp\. 125–136\.Cited by:[§1](https://arxiv.org/html/2605.13438#S1.p3.1),[§2\.2](https://arxiv.org/html/2605.13438#S2.SS2.p1.1),[§5](https://arxiv.org/html/2605.13438#S5.p2.1)\.
- L\. Zheng, W\. Chiang, Y\. Sheng, S\. Zhuang, Z\. Wu, Y\. Zhuang, Z\. Lin, Z\. Li, D\. Li, E\. Xing,et al\.\(2023\)Judging llm\-as\-a\-judge with mt\-bench and chatbot arena\.Advances in neural information processing systems36,pp\. 46595–46623\.Cited by:[Table 4](https://arxiv.org/html/2605.13438#S4.T4),[Table 4](https://arxiv.org/html/2605.13438#S4.T4.9.2.1)\.
- W\. Zhong, L\. Guo, Q\. Gao, H\. Ye, and Y\. Wang \(2024\)Memorybank: enhancing large language models with long\-term memory\.InProceedings of the AAAI conference on artificial intelligence,Vol\.38,pp\. 19724–19731\.Cited by:[item 3](https://arxiv.org/html/2605.13438#S3.I1.i3.p1.1),[§3\.1](https://arxiv.org/html/2605.13438#S3.SS1.p4.1)\.

## Appendix ARelated Work

### A\.1Graph\-based Agent Memory

Current graph\-augmented memory systems generally evolve along three axes, all of which treat the graph as a query\-time artifact rather than a living substrate\. HippoRAG\[Gutiérrezet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib33),[2025](https://arxiv.org/html/2605.13438#bib.bib34)\]maps memory to a static knowledge graph retrieved via Personalized PageRank\. However, this structurally frozen approach struggles to handle the evolving lifecycle of agent interactions\. For dynamic updates, Mem0\[Chhikaraet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib43)\]performs per\-turn memory management through LLM\-driven text rewrites\. MAGMA\[Jianget al\.,[2026](https://arxiv.org/html/2605.13438#bib.bib36)\]decomposes memory into multiple orthogonal graphs and supplements them with an LLM\-driven slow\-path inference pass\. Zep\[Rasmussenet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib68)\]introduces bi\-temporal validity to track fact invalidation\. Across these paradigms, although several systems do process events incrementally, the graph’stopologygrows by accumulation, rewriting, or invalidation, but does notmetabolisebetween events—it does not fold redundant fragments, decay stale connections by mere passage of time, or reconnect orphans through associative similarity\. Real\-time topological metabolism under an event stream therefore remains largely unaddressed\. Table[6](https://arxiv.org/html/2605.13438#A1.T6)situates representative systems—together with the non\-graph alternatives discussed below—on four orthogonal axes \(stream input, proactive update, evolving topology, symbolic and inspectable\), making the absent corner explicit\.

Table 6:Design\-space for agent memory under continuous streams\.Four orthogonal axes distinguish always\-on proactive memory from reactive, batch\-oriented architectures;CogniFoldis the only system satisfying all four simultaneously\. Table[12](https://arxiv.org/html/2605.13438#A3.T12)provides the complementary engineering\-coverage view \(which of the four structural debts of §[3\.2](https://arxiv.org/html/2605.13438#S3.SS2)each system addresses\)\. ✓ = satisfies; — = does not;*partial*= subset of the axis\.
### A\.2Non\-Topological Test\-Time Learning

A parallel line of work pursues persistent test\-time learning through non\-graph mechanisms, revealing a trade\-off between inspectability and structural depth\. Generative Agents\[Parket al\.,[2023](https://arxiv.org/html/2605.13438#bib.bib9)\]and MemGPT\[Packeret al\.,[2023](https://arxiv.org/html/2605.13438#bib.bib11)\]treat memory as natural language to be paged or searched, while systems like A\-Mem\[Xuet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib35)\]rewrite Zettelkasten\-style notes upon receiving new information\. These text\-rewriting methods remain highly inspectable but are structurally blind, as the underlying architecture does not evolve\. Conversely, Titans\[Behrouzet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib42)\]introduces a neural long\-term memory module updated via surprise\-driven gradient descent during inference\. While computationally efficient, this implicit approach deposits knowledge into an opaque weight space that cannot be audited or selectively deleted\[McCloskey and Cohen,[1989](https://arxiv.org/html/2605.13438#bib.bib44)\]\. Consequently, the field remains divided between methods that are transparent but topologically static, and those that learn continuously but sacrifice discrete, mutable geometry\.

### A\.3Proactive Agent Memory

Most agent\-memory systems are fundamentally designed for the reactive setting: the user issues a query, and the system retrieves context to generate a response\. Explicit treatment of the always\-on setting—where input is a continuously arriving event stream—remains sparse\. Traditional BDI\-style architectures\[Bratman,[1987](https://arxiv.org/html/2605.13438#bib.bib54)\]address agency by deriving intentions from hardcoded rules and explicit goals, rather than letting them emerge bottom\-up from accumulated evidence\. Recent prospective\-memory work\[Einstein and McDaniel,[2005](https://arxiv.org/html/2605.13438#bib.bib4)\]and cognitive\-agent frameworks\[Sumerset al\.,[2023](https://arxiv.org/html/2605.13438#bib.bib12), Shinnet al\.,[2023](https://arxiv.org/html/2605.13438#bib.bib10)\]acknowledge the need for goal\-directed organization but rarely anchor intent emergence to the evolving topology of the memory itself\. Even with advanced read\-path innovations, mainstream systems assume a stable graph at query time without defining how the structure should reorganize between queries, leaving a critical gap in anticipatory, autonomous memory assembly\.

## Appendix BHyperparameters

All numeric defaults referenced in §[3\.1](https://arxiv.org/html/2605.13438#S3.SS1)are listed here for reproducibility\. Values were tuned empirically on a held\-out subset of the personal\-timeline simulator and held fixed across all reported experiments\.

### B\.1Default Edge Weights \(Table[1](https://arxiv.org/html/2605.13438#S2.T1)\)

Table 7:Prior weights at edge creation\. Per\-instance weights are then updated dynamically throughReinforcesstrengthening and exponential edge decay \(§[3\.2](https://arxiv.org/html/2605.13438#S3.SS2)\)\.
### B\.2Write\-Path Scoring \(Eq\.[1](https://arxiv.org/html/2605.13438#S3.E1)\)

Table 8:Write\-path scoring weights and context\-window allocation\. Tier sub\-weights:*Immediate*—70% recency \+ 30% urgency;*Working*—50% PageRank \+ 30% recency \+ 20% type;*Background*—80% PageRank \+ 20% diversity\.
### B\.3Consolidation Operations \(§[3\.2](https://arxiv.org/html/2605.13438#S3.SS2)\)

Table 9:Consolidation operations that discharge the four structural debts\. Default weights for each typed edge are listed in Table[1](https://arxiv.org/html/2605.13438#S2.T1)\.
### B\.4Intent Emergence \(Eq\.[2](https://arxiv.org/html/2605.13438#S3.E2)\)

Table 10:Per\-category EMA loop for adaptive intent\-emission threshold\.Table 11:Read\-path retrieval and traversal parameters\.

## Appendix CFour\-Debt Attack Surface

Table 12:Four\-debt attack\-surface comparison\.An event\-stream memory graph accumulates four structural debts by the nature of its input \(Section[3\.2](https://arxiv.org/html/2605.13438#S3.SS2)\):*accumulation*,*compression*,*decay*,*completion*\. “—” = not addressed; “LLM rewrite” = text\-layer rewrite only; “partial” = subset of the debt; ✓= automatic graph\-level operation\.Scope of comparison: all systems with a published description of their ingestion and update mechanisms in the 2024–2026 graph\-memory literature; closed industrial systems without published mechanisms are excluded\.CogniFoldis the only system to address all four debts as automatic graph\-level operations\.†\\daggerZep/Graphiti provides bi\-temporal fact invalidation \(t\_invalid\) on*contradiction*, which addresses*consistency*rather than*decay*: no edge weakens with passage of time alone\. Consistency is orthogonal to the four debts in our taxonomy\.

Table[12](https://arxiv.org/html/2605.13438#A3.T12)compares how each prior agent\-memory system addresses \(or fails to address\) the four structural debts of §[3\.2](https://arxiv.org/html/2605.13438#S3.SS2)—accumulation, compression, decay, and completion\.CogniFoldis the first system to address all four as automatic, topology\-level operations\.

## Appendix DImplementation

Python 3\.11: NetworkX \(graph\), Pydantic v2 \(schemas\), LangGraph \(agent\), FastAPI \(HTTP\), FAISS \(ANN index\)\. 21 packages, 1,167 tests, strict typing\. Deployed on GCP Cloud Run with per\-session isolation\. Multi\-domain support \(learning, finance, programming\) via prompt profiles\.

## Appendix ECore System Prompt \(Excerpt\)

The LLM agent receives a system prompt composed from modular sections\. The key section governing concept extraction:

Youareacognitivegraphagent\.Givenanevent

andcontext,produceanUpdatePlanwithoperations:

NODETYPES:event\(rawinput\),concept\(patterns\),

intent\(goals\),time\(deadlines\)

EDGETYPESwithweights:

GROUNDS\(0\.9\):event\-\>concept/intent

REINFORCES\(0\.7\):event\-\>existingconcept

TRIGGERS\(0\.8\):concept\-\>intent

PART\_OF\(0\.7\):concept\-\>concept\(hierarchy\)

RULES:

1\.Createconceptsforrecurringpatterns\(3\+events\)

2\.Linkeveryconcepttogroundingevents

3\.Mergenear\-duplicateconceptsviaMERGE\_NODES

4\.Createintentsonlywhenpatternssuggest

unmetgoalswithsupportingevidence

5\.Self\-review:checkformissingedgesbetween

conceptsthatsharegroundingevents

The full prompt includes 20 composable sections \(edge types, connectivity rules, validation checklist, deduplication, self\-review\)\. Domain\-specific YAML profiles override the role section while retaining structural sections\.

## Appendix FCogEval\-Bench Details

### F\.1Gold Graph Schemas

For each scenario we manually define a gold concept graph𝒢∗=\(𝒞∗,ℛ∗,ℋ∗,ℐ∗\)\\mathcal\{G\}^\{\*\}=\(\\mathcal\{C\}^\{\*\},\\mathcal\{R\}^\{\*\},\\mathcal\{H\}^\{\*\},\\mathcal\{I\}^\{\*\}\)with four components\.Concepts𝒞∗\\mathcal\{C\}^\{\*\}: 8–9 non\-hierarchical concepts, each specified with label, natural\-language description, representative keywords, and expected event count\. Concepts are grounded in established domain knowledge \(e\.g\., SoftEng includes Code Review, Sprint Planning, Deployment\) and deliberately avoidCogniFold\-specific abstractions to keep the gold standard system\-agnostic\.Relationshipsℛ∗\\mathcal\{R\}^\{\*\}: 9–14 labelled inter\-concept edges using four relationship types drawn from cognitive schema theory\[Gilboa and Marlatte,[2017](https://arxiv.org/html/2605.13438#bib.bib6), Bartlett,[1995](https://arxiv.org/html/2605.13438#bib.bib7)\]:PART\_OF\(compositional hierarchy\),TRIGGERS\(temporal causation\),REINFORCES\(feedback strengthening\),CAUSES\(interventional causation\)\.Hierarchy parentsℋ∗\\mathcal\{H\}^\{\*\}: 1–3 superordinate concepts \(e\.g\., Work Projects subsumes Coding Sessions and Code Review\) that test hierarchical abstraction\.Expected intentsℐ∗\\mathcal\{I\}^\{\*\}: 2 goal nodes per scenario, each grounded in 2–3 supporting concepts with a specified trigger pattern \(e\.g\., “3\+ exercise events within the week”→\\toMaintain Regular Exercise\)\. Each gold graph additionally contains 2planted multi\-hop reasoning chains\(3–4 hops\) following the compositional methodology of MuSiQue\[Trivediet al\.,[2022](https://arxiv.org/html/2605.13438#bib.bib22)\]: chains are sequences of events connected through causal or temporal links across different concepts \(e\.g\., deployment→\\tostaging bug→\\toclient demo failure\), requiring cross\-concept traversal to reconstruct\. Complete gold concept graphs for all 6 scenarios are provided in the supplementary material \(benchmarks/cogeval/data/gold\_graphs/\)\.

### F\.2Event Generation Pipeline

From each gold graph, a grounded event stream is generated in four stages\. \(1\)Concept\-grounded events: for each conceptc∈𝒞∗c\\in\\mathcal\{C\}^\{\*\}, GPT\-4o\-mini generatesncn\_\{c\}realistic first\-person events conditioned on the concept’s label, description, and keywords; each event receives the by\-construction labelgold\_concept=c\\text\{gold\\\_concept\}=c\. \(2\)Chain events: for each planted chain, the pipeline generates sequential events following the chain’s step descriptions, ensuring entity consistency across hops \(e\.g\., the same “auth service v2\.3” appears in the deployment, bug discovery, and client demo events\)\. \(3\)Distractor injection: 10–15% of events come from unrelated topics \(weather, unrelated news, etc\.\), labelledgold\_concept=null\\text\{gold\\\_concept\}=\\texttt\{null\}\. \(4\)Temporal shuffling: all events receive timestamps spanning the scenario’s temporal window \(5–60 days\), then are sorted chronologically with chain events interleaved among non\-chain events to prevent trivial sequential pattern matching\.

### F\.3Scenarios

Table 13:CogEval\-Bench scenarios\.Six domains with controlled gold concept graphs\. Each scenario defines gold concepts, planted multi\-hop chains \(3–4 hops\), expected intents, and distractor events \(∼15%\{\\sim\}15\\%\)\. Events generated by GPT\-4o\-mini from gold graphs, temporally shuffled\.Six scenarios span four domains \(Table[13](https://arxiv.org/html/2605.13438#A6.T13)\): professional work \(SoftEng\), medical recovery \(Health\), team coordination \(Team\), breaking news \(News\), academic research \(Academic\), and customer support \(Support\)\. These domains test distinct cognitive patterns: daily routines with gradual concept consolidation, crisis cascades requiring causal chain tracking, topic drift across independent threads, deep abstraction hierarchies, and repetitive pattern detection\. The benchmark totals 251 events, 49 gold concepts, 12 planted multi\-hop chains, and 12 expected intents across the six scenarios\.

### F\.4Evaluation Metrics

Three tracks are computed per system and averaged across scenarios\.Track A: Concept Emergence—\(i\)Gold F1: precision and recall of system concepts against gold concepts via embedding\-based soft matching \(text\-embedding\-3\-small, cosine≥0\.75\\geq 0\.75\) with optimal one\-to\-one assignment via the Hungarian algorithm\[Kuhn,[1955](https://arxiv.org/html/2605.13438#bib.bib57)\]; \(ii\)LLM Quality: GPT\-4o\-mini judge rates meaningfulness, groundedness, and abstraction level, each 0–1, then averaged; \(iii\)Harmony: harmonic mean of Gold F1 and LLM Quality; \(iv\)Purity: average pairwise embedding similarity among events grounding each concept\.Track B: Relationship Topology—Chain Discovery Rate \(fraction of planted chains recoverable via BFS between endpoints\), Clustering Coefficient\[Newman,[2006](https://arxiv.org/html/2605.13438#bib.bib39)\], Modularity \(Newman Q\), Edge Type Entropy\.Track C: Compression & Proactivity—Compression Ratio \(input events / output concepts\), PageRank Gini, Proactivity \(fraction of intents with≥2\\geq 2grounding connections\)\.

### F\.5Comparison Systems

Seven systems under identical LLM and events\.OpenIE KG: HippoRAG\-style triples, flat entity graph, no concept folding\.Cognee\[Topoteretes,[2026](https://arxiv.org/html/2605.13438#bib.bib41)\]: Extract–Cognify–Load pipeline building property graphs via LLM entity extraction and classification; no cross\-event merging or temporal tracking\.HippoRAG 2\[Gutiérrezet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib34)\]: deeper passage integration, synonym expansion edges, PPR retrieval; still entity\-level with no concept abstraction\.GraphRAG\[Edgeet al\.,[2024](https://arxiv.org/html/2605.13438#bib.bib40)\]: LLM entity/relation extraction, Leiden community detection, LLM summarization per community; batch, no temporal folding or intent emergence\.Mem0\[Chhikaraet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib43)\]: text\-rewrite memory cells with vector retrieval; no native graph \(we materialise each cell as a node and induce vector\-similarity edges\)\.Zep\[Rasmussenet al\.,[2025](https://arxiv.org/html/2605.13438#bib.bib68)\]: temporal knowledge\-graph memory with entity\-centric edges and time stamps\.CogniFold: full merge\-fold pipeline \(events→\\toconcepts→\\tointents\), 9 typed edges, online event\-by\-event processing, temporal decay, lifecycle management\.

### F\.6Per\-Scenario Results

Table 14:CogEval\-Bench per\-scenario Harmony scores\.Harmony = harmonic mean of gold F1 and LLM quality \(Track A\)\.CogniFoldconsistently leads across all 6 scenarios despite their diversity\. Mem0’s flat memory store \(no concept abstraction\) yields zero across the board; entity\-level systems \(OpenIE KG, Cognee, HippoRAG 2, Zep/Graphiti\) cluster around 0\.07–0\.19; community\-level GraphRAG reaches 0\.23–0\.39;CogniFold’s folding achieves 0\.38–0\.57\.KG = OpenIE KG; Cog = Cognee; HR2 = HippoRAG 2; GR = GraphRAG; M0 = Mem0; Zp = Zep/Graphiti; CF =CogniFold\. Compression = events/concepts \(×\\times\)\. Clustering markedboldforCogniFolddue to semantic meaningfulness; HippoRAG 2’s high clustering arises from synonym edges, Zep’s from dense entity co\-mention edges, and Mem0’s from undifferentiated similarity\-induced edges \(see text\)\.

Table[14](https://arxiv.org/html/2605.13438#A6.T14)provides per\-scenario breakdowns on three key metrics\.CogniFold’s Harmony ranges from 0\.383 \(Team\) to 0\.572 \(SoftEng\), consistently above GraphRAG’s best \(0\.390, Support\)\. Compression is robust \(3\.4–5\.5×\\times\), highest on Support and Team where repetitive event patterns benefit most from folding\. Clustering ranges from 0\.132 \(Health\) to 0\.392 \(Support\), lowest on Health where longer temporal spans create sparser event overlap\.

### F\.7Why Enrichment≠\\neqAbstraction

Cognee and HippoRAG 2 occupy the same Harmony tier as vanilla OpenIE KG despite adding substantial computational machinery \(ECL pipeline, synonym expansion, Personalized PageRank\)\. This confirms that the bottleneck in concept emergence is not extraction depth butcross\-event integration—the ability to recognize that eventse1,e3,e7e\_\{1\},e\_\{3\},e\_\{7\}ground the same underlying concept and merge them into a single abstraction\. Systems that process each event in isolation, however thoroughly, cannot produce this integration\.

### F\.8Proactivity as Emergence Evidence

Proactivity captures a property no baseline can produce: identifying goals from converging evidence before being asked\. An average of 22 intent nodes per scenario, with 61\.4% well\-grounded by multiple events, constitutes evidence of intent emergence beyond memory organization\. While these intents are LLM\-generated rather than autonomously emergent from prediction error, they represent a measurable step toward proactive intelligence that flat memory systems cannot take\.

### F\.9Construct Validity

Two metrics that most sharply separateCogniFoldfrom baselines—Purity \(event\-level grounding\) and Proactivity \(intents with≥2\\geq 2groundings\)—measure representational features that onlyCogniFoldemits:GROUNDSedges for Purity, intent nodes for Proactivity\. A reader may reasonably worry the metric is biased toward our representation\. We address this risk with three choices: \(i\) the representational hierarchy holds not only on these two metrics but also on task\-agnostic ones \(Harmony, Gold F1, LLM Quality, Compression, Clustering\), on which baselines are free to score well; \(ii\) CogEval is deliberately supplemented by the broad 7\-benchmark memory results \(§[4\.5](https://arxiv.org/html/2605.13438#S4.SS5)\), which use external benchmarks and metrics; \(iii\) we report raw intent counts and grounding densities, not normalized scores\. Even so, a fully adversarial benchmark—one where our representation could lose—is future work\.

### F\.10Scale and Scope Limitations

CogEval\-Bench uses synthetic events generated from predefined gold graphs, which may not capture the full complexity of natural event streams\. The scale is deliberately small \(∼42\{\\sim\}42events per scenario\) to enable controlled structural evaluation; whether the structural hierarchy holds at larger scales remains to be validated\. The LLM judge introduces potential evaluation bias; we mitigate this by per\-concept independent scoring \(not pairwise comparison\) and report the full judge prompt above\.

### F\.11Event Generation Prompt

For each conceptccwith labelℓ\\elland descriptiondd, events are generated using the following prompt template:

Generate\{n\}realisticfirst\-personeventsfora

\{scenario\}scenario,groundedintheconcept

"\{label\}":\{description\}\.

Eacheventshould:

\-Beaspecific,timestampedexperience\(notgeneric\)

\-Usefirst\-personperspective

\-Beself\-contained\(understandablewithoutcontext\)

\-Relateclearlytotheconceptkeywords:\{keywords\}

OutputasJSONarraywithfields:title,description,

timestamp,event\_type\.

Chain events use a sequential prompt that references entities from prior chain steps to maintain cross\-hop consistency\. Distractor events are generated from unrelated topics \(e\.g\., weather, unrelated hobbies\) without concept grounding\.

### F\.12LLM Judge Prompt

The concept quality judge evaluates each system concept independently:

Evaluatethisconceptextractedfromaneventstream\.

Concept:"\{concept\_label\}"

Groundingevents\(ifany\):\{event\_summaries\}

Scenariocontext:\{scenario\_description\}

Rateonthreedimensions\(0\.0to1\.0\):

1\.MEANINGFULNESS:Isthisasemanticallycoherent

concept?\(1\.0=clearlydefinedtheme/pattern,

0\.0=incoherentortrivial\)

2\.GROUNDEDNESS:Istheconceptwell\-supportedby

itsgroundingevents?\(1\.0=strongevidence,

0\.0=nosupportingevidence\)

3\.ABSTRACTIONLEVEL:Isthistherightlevelof

abstraction?\(1\.0=usefulgeneralization,

0\.0=toospecificortoovague\)

OutputJSON:\{"meaningfulness":X,"groundedness":X,

"abstraction":X\}

### F\.13Embedding Similarity Threshold Sensitivity

The Gold F1 computation uses cosine similarity≥0\.75\\geq 0\.75for concept matching\. Table[15](https://arxiv.org/html/2605.13438#A6.T15)reportsCogniFold’s Harmony score averaged across 6 scenarios at varying thresholds, confirming that the ranking is robust to threshold choice\.

Table 15:Sensitivity of Harmony to embedding similarity threshold \(CogniFold, averaged over 6 scenarios\)\.Higher thresholds are more conservative \(fewer matches, lower recall\); lower thresholds are more permissive \(more matches, risk of false positives\)\. The relative ranking of systems is preserved across all tested thresholds\.

## Appendix GReproducibility

Code and data\.Source code, benchmark runner scripts, and evaluation harnesses will be released upon publication\. All benchmarks use publicly available datasets: MuTualCuiet al\.\[[2020](https://arxiv.org/html/2605.13438#bib.bib21)\], ToMiLeet al\.\[[2019](https://arxiv.org/html/2605.13438#bib.bib97)\], MuSiQueTrivediet al\.\[[2022](https://arxiv.org/html/2605.13438#bib.bib22)\], NarrativeQAKočiskỳet al\.\[[2018](https://arxiv.org/html/2605.13438#bib.bib32)\], StreamingQALiskaet al\.\[[2022](https://arxiv.org/html/2605.13438#bib.bib26)\], LoCoMoMaharanaet al\.\[[2024](https://arxiv.org/html/2605.13438#bib.bib63)\], and BABILongKuratovet al\.\[[2024](https://arxiv.org/html/2605.13438#bib.bib20)\]\.

Compute\.All experiments use the OpenAI API \(GPT\-4o\-mini for generation, text\-embedding\-3\-small for embeddings\)\. Graph construction and retrieval run on a single CPU; no GPU is required\. Total API cost for all experiments \(including development iterations\): approximately $150\.

Randomness\.The primary source of non\-determinism is LLM sampling \(temperature 0\.0 for all evaluation calls\)\. Graph construction is deterministic given the same LLM outputs\. Benchmark sampling uses fixed random seeds \(seed=42 for all dataset splits\)\. Wilson confidence intervals are reported for full\-scale results \(n≥100n\{\\geq\}100\)\.

Broader impact\.CogniFoldis a general\-purpose cognitive memory architecture\. Like any persistent memory system, it raises privacy considerations: the graph accumulates personal information that persists across sessions\. Production deployments should implement access controls and data retention policies\. The system does not autonomously take actions; intent nodes represent identified goals, not executed plans\.