ConvMemory v3: A Validity Context Layer for Conversational Memory via Target-Conditioned Relation Verification

# A Validity Context Layer for Conversational Memory via Target-Conditioned Relation Verification
Source: [https://arxiv.org/html/2606.26753](https://arxiv.org/html/2606.26753)
Taiheng Pan School of Computing and Information Systems University of Melbourne github\.com/pth2002

\(June 2026\)

###### Abstract

Conversational memory retrieval optimizes relevance, yet a retrieved memory can be relevant and simultaneously outdated: a later turn updates, corrects, or supersedes it\. ConvMemory v3 adds a*validity context layer*that detects and surfaces this update evidence through target\-conditioned relation verification, sitting after the v1/v2 retrieval path\. The core mechanism is a dual\-evidence gate that conditions a relation judgment on the specific target proposition, scoring a \(target, source\) pair through the product of a MiniLM slot head and a DeBERTa\-v3 slot head and gating it by conservative event/operation evidence\. On a synthetic multi\-hop validity benchmark the gate reaches90\.12%±1\.7390\.12\\%\\pm 1\.73accuracy; through a real\-data feedback loop that mines failure patterns but trains on synthetic pairs only, the verifier transfers to Memora role binding with zero target\-side labels, reaching98\.8%±0\.998\.8\\%\\pm 0\.9group\-all\-correct\. The deployed layer preserves retrieval by default: acontextmode attaches structured validity metadata while keeping the candidate set and rank order fixed, and a query\-conditioneddemotemode is an explicit opt\-in for dense current\-state workloads, where it raises current\-active H@1 from a never\-demote baseline of45\.1%45\.1\\%to95\.7%±1\.295\.7\\%\\pm 1\.2while protecting non\-superseded memories at99\.4%99\.4\\%recall\. Six machine\-verifiable safety contracts pin the layer’s behavior\. Multi\-hop graph propagation is validated as a mechanism; fully automatic construction of strict prerequisite edges is characterized as a boundary, since strict necessity requires counterfactual world knowledge\. This report extends ConvMemory v1\[[1](https://arxiv.org/html/2606.26753#bib.bib1)\]and v2\[[2](https://arxiv.org/html/2606.26753#bib.bib2)\]\.

###### Contents

1. [1Introduction](https://arxiv.org/html/2606.26753#S1)
2. [2Relationship to the v1 and v2 Reports](https://arxiv.org/html/2606.26753#S2)
3. [3Related Work](https://arxiv.org/html/2606.26753#S3)
4. [4The Validity Problem](https://arxiv.org/html/2606.26753#S4)
5. [5The Target\-Conditioned Dual\-Evidence Gate](https://arxiv.org/html/2606.26753#S5)1. [5\.1Slot heads](https://arxiv.org/html/2606.26753#S5.SS1) 2. [5\.2Conservative event/operation gate](https://arxiv.org/html/2606.26753#S5.SS2) 3. [5\.3Aggregation by noisy\-or](https://arxiv.org/html/2606.26753#S5.SS3) 4. [5\.4A worked example](https://arxiv.org/html/2606.26753#S5.SS4) 5. [5\.5Architecture freeze: head merge and the load\-bearing spine](https://arxiv.org/html/2606.26753#S5.SS5)
6. [6Training the Verifier](https://arxiv.org/html/2606.26753#S6)1. [6\.1Objective](https://arxiv.org/html/2606.26753#S6.SS1) 2. [6\.2Stratified supervision](https://arxiv.org/html/2606.26753#S6.SS2) 3. [6\.3The real\-data feedback loop](https://arxiv.org/html/2606.26753#S6.SS3) 4. [6\.4The deployed query\-conditioned calibrator](https://arxiv.org/html/2606.26753#S6.SS4)
7. [7Experimental Protocol](https://arxiv.org/html/2606.26753#S7)
8. [8Verifier Progression](https://arxiv.org/html/2606.26753#S8)
9. [9Role\-Binding Verification](https://arxiv.org/html/2606.26753#S9)
10. [10Multi\-Hop Validity Propagation](https://arxiv.org/html/2606.26753#S10)
11. [11The Validity Context Layer in Dense Retrieval](https://arxiv.org/html/2606.26753#S11)
12. [12LoCoMo Density Audit](https://arxiv.org/html/2606.26753#S12)
13. [13Cost\-Conscious Routing](https://arxiv.org/html/2606.26753#S13)
14. [14Deployment: Modes, Contracts, and Cost](https://arxiv.org/html/2606.26753#S14)1. [14\.1Modes](https://arxiv.org/html/2606.26753#S14.SS1) 2. [14\.2Safety contracts](https://arxiv.org/html/2606.26753#S14.SS2) 3. [14\.3Cost and configuration](https://arxiv.org/html/2606.26753#S14.SS3) 4. [14\.4Released checkpoint figures](https://arxiv.org/html/2606.26753#S14.SS4)
15. [15Boundary Findings](https://arxiv.org/html/2606.26753#S15)
16. [16Reproducibility](https://arxiv.org/html/2606.26753#S16)
17. [17Discussion](https://arxiv.org/html/2606.26753#S17)1. [17\.1Limitations](https://arxiv.org/html/2606.26753#S17.SS1) 2. [17\.2Future Work](https://arxiv.org/html/2606.26753#S17.SS2)
18. [ASource\-of\-Truth Checklist](https://arxiv.org/html/2606.26753#A1)
19. [BReleased Checkpoint Provenance](https://arxiv.org/html/2606.26753#A2)
20. [References](https://arxiv.org/html/2606.26753#bib)

## 1Introduction

A conversational agent accumulates memories across many sessions, and a query against that store asks not only “which memory is relevant” but increasingly “which memory is still true\.” A user who changes jobs, moves cities, or revises a preference leaves a trail in which an earlier memory remains topically relevant while a later memory supersedes it\. ConvMemory v1\[[1](https://arxiv.org/html/2606.26753#bib.bib1)\]and v2\[[2](https://arxiv.org/html/2606.26753#bib.bib2)\]both optimize relevance: v1 organizes a high\-recall top\-500 pool with a lightweight learned reranker, and v2 adds a recall\-preserving top\-10 evidence reranker that improves ordering inside the protected set\. Neither layer reasons about whether a relevant memory has been overturned\.

ConvMemory v3 adds that reasoning as a*validity context layer*\. The layer runs after the v1/v2 retrieval path and answers a different question for each retrieved memory: relative to the specific target the query is about, does a later memory update, correct, or supersede this one? The answer is attached as structured validity metadata, and the agent decides how to use it\. By default thecontextmode preserves the candidate set and rank order while extending the result schema with validity metadata; the v1/v2 retrieval output is unchanged, and theoff/Nonemode is byte\-identical to a build without v3\.

The technical core is a*target\-conditioned relation verifier*\. A relation judgment that ignores the target collapses on the cases that matter: two memories about the same person can stand in an update relation for one attribute and an unrelated relation for another, and the relation is only well\-defined once the target proposition is fixed\. v3 conditions the judgment on the target through a dual\-evidence gate that multiplies a MiniLM slot score by a DeBERTa\-v3 slot score, gates the product by conservative event/operation evidence, and propagates the result over sources by noisy\-or\. A stratified\-supervision feedback loop learns from real role\-binding hard cases and transfers back with zero target\-side labels\.

We make three contributions\.

1. 1\.A target\-conditioned dual\-evidence gate\(the mechanism\): a slot\-product verifier with a conservative event/operation gate that reaches90\.12%±1\.7390\.12\\%\\pm 1\.73on a synthetic multi\-hop validity benchmark and, through a real\-data feedback loop, transfers to Memora role binding at98\.8%±0\.998\.8\\%\\pm 0\.9group\-all\-correct with zero target\-side labels\. This is well above a target\-position rule \(78\.6%78\.6\\%\), zero\-shot NLI \(64\.2%64\.2\\%\), and a relevance cross\-encoder \(17\.9%17\.9\\%\)\.
2. 2\.A validity context layer\(the system\): an opt\-in stage exposing four accepted values through a singlevalidity\_modeargument \(None,off,context,demote\), corresponding to three behavioral modes\. Thecontextdefault preserves the candidate set and rank order while attaching validity evidence; thedemotemode is an explicit opt\-in that reorders for dense current\-state workloads, raising current\-active H@1 from45\.1%45\.1\\%to95\.7%±1\.295\.7\\%\\pm 1\.2on Memora dense retrieval while protecting non\-superseded memories at99\.4%99\.4\\%recall\. Six machine\-verifiable safety contracts pin the layer’s behavior, backed by 41 passing tests\.
3. 3\.Two boundary findings\(the scope\): multi\-hop graph propagation is validated as a mechanism with given structure; target\-conditioned relation labels for free\-form multi\-hop conversation are, to our knowledge, scarce, and fully automatic construction of strict prerequisite edges requires counterfactual necessity judgment that a discriminative model does not provide reliably\. Both are characterized precisely in §[15](https://arxiv.org/html/2606.26753#S15)\.

v3 builds directly on the ConvMemory cascade: v1 supplies the candidate pool, v2 supplies the protected ordering, and v3 supplies validity context\. Two models carry distinct roles, and the report keeps them separate throughout\. The dual\-evidence verifier establishes that target\-conditioned relation verification works, and supplies the synthetic and role\-binding evidence for the mechanism\. The deployed demotion calibrator is the query\-conditioned model the released package runs, and it produces the dense\-retrieval current\-active H@1 of95\.7%95\.7\\%\. The verifier, its training, the propagation mechanism, and the deployed calibrator are described in turn, each with the experiments that establish it\.

## 2Relationship to the v1 and v2 Reports

ConvMemory v1\[[1](https://arxiv.org/html/2606.26753#bib.bib1)\]introduced a lightweight learned memory reranker, a negative attribution result on a learned temporal window, and a research\-preview conflict editor\. ConvMemory v2\[[2](https://arxiv.org/html/2606.26753#bib.bib2)\]introduced a recall\-preserving top\-10 evidence reranker that reorders only v1’s protected top\-10 and improves MRR and H@1 while keeping Recall@10 fixed by construction\. v3 composes with both rather than replacing either\.

#### v1 \(the conflict\-editor research preview\)\.

v1 shipped a research\-preview conflict editor and listed conflict modeling as open\. v3 is the matured successor to that direction: instead of editing conflicting memories in place, v3 verifies the relation between a target memory and a later source and surfaces the result as validity context\. The v3 verifier is target\-conditioned and supervised, where the v1 preview was neither\.

#### v2 \(the recall\-preserving cascade\)\.

v2 reorders v1’s protected top\-10 and leaves recall fixed\. v3 attaches after that cascade\. In thecontextdefault, v3 preserves whatever ordering v1/v2 produced and only adds validity metadata, so a v1\+v2\+v3 deployment retains v2’s exact retrieval behavior\. Thedemoteopt\-in is the single mode in which v3 reorders, and it is scoped to dense current\-state workloads \(§[11](https://arxiv.org/html/2606.26753#S11)\)\.

#### The temporal\-window negative result\.

v1 reported that a learned temporal window was statistically significant on aggregate but not temporally specific, and v2 made no temporal\-mechanism assertion\. v3’s mechanism is target\-conditioned relation verification on memory text, distinct from a temporal window: the validity signal comes from what two memories assert about a shared target, not from their position in time\. The v1 negative result stands unchanged\.

## 3Related Work

v3 sits at the intersection of three lines of work, and its contribution is clearest when contrasted with each\.

#### Agent memory systems\.

A growing body of systems give conversational agents persistent memory\. Systems such as MemGPT\[[11](https://arxiv.org/html/2606.26753#bib.bib11)\], Generative Agents\[[12](https://arxiv.org/html/2606.26753#bib.bib12)\], MemoryBank\[[13](https://arxiv.org/html/2606.26753#bib.bib13)\], and Mem0\[[14](https://arxiv.org/html/2606.26753#bib.bib14)\]write summaries or facts to a store, index them, and retrieve them by relevance when a later query needs history\. Their objective is coverage and recall, making sure the right memory can be found again\. v3 is orthogonal and complementary: it does not change what is stored or how it is retrieved, but adds a layer that judges, for a retrieved memory, whether a later memory has superseded it with respect to the target the query asks about\. Storage and retrieval answer “can this memory be found”; the validity layer answers “is this memory still true\.” Benchmarks for long\-term conversational memory, such as LoCoMo\[[16](https://arxiv.org/html/2606.26753#bib.bib16)\]and LongMemEval\[[17](https://arxiv.org/html/2606.26753#bib.bib17)\], mostly frame the task as fact retrieval from past sessions; Memora\[[15](https://arxiv.org/html/2606.26753#bib.bib15)\]is closest to v3 in foregrounding memory mutation and obsolescence, which is the regime the validity layer targets\.

#### Knowledge editing\.

Knowledge\-editing methods such as ROME\[[9](https://arxiv.org/html/2606.26753#bib.bib9)\]and MEMIT\[[10](https://arxiv.org/html/2606.26753#bib.bib10)\]update what a model knows by editing its weights, so that the model itself stops asserting an outdated fact\. v3 leaves model weights untouched and operates on the retrieval context instead: it annotates or reorders the retrieved memories that a downstream model will read\. Editing changes the model; the validity layer changes the evidence the model is given, which keeps the original memories auditable and the layer removable\.

#### Natural language inference and fact verification\.

NLI and fact\-verification models, trained on datasets such as FEVER\[[8](https://arxiv.org/html/2606.26753#bib.bib8)\], judge whether one sentence entails, contradicts, or supports another\. v3 uses an NLI backbone for one of its slot heads, but the task is different: fact verification asks whether two sentences are consistent in isolation, whereas the validity layer asks whether a source overturns a target*with respect to a specific proposition*\. A sentence\-level contradiction signal that ignores the target conflates an employer update with an unrelated hometown statement; the target\-conditioned slot product is precisely what a query\-blind NLI judgment lacks \(§[11](https://arxiv.org/html/2606.26753#S11)\)\.

The components v3 builds on are concrete\. The verifier uses two slot heads: a MiniLM cross\-encoder\[[4](https://arxiv.org/html/2606.26753#bib.bib4)\]111cross\-encoder/ms\-marco\-MiniLM\-L6\-v2\.of the family v1 used as a distillation teacher and v2 used as its evidence scorer, fine\-tuned on MS MARCO passage ranking\[[7](https://arxiv.org/html/2606.26753#bib.bib7),[6](https://arxiv.org/html/2606.26753#bib.bib6)\], and a DeBERTa\-v3 backbone\[[5](https://arxiv.org/html/2606.26753#bib.bib5)\]initialized from an NLI checkpoint222cross\-encoder/nli\-deberta\-v3\-base, based onmicrosoft/deberta\-v3\-baseand trained on SNLI and MultiNLI\.and fine\-tuned with a binary classification head\. The retrieval path beneath v3 is the ConvMemory v1 reranker over a dense MPNet\[[3](https://arxiv.org/html/2606.26753#bib.bib3)\]top\-500 pool, optionally followed by the v2 evidence reranker\. Evaluation uses three settings: a synthetic multi\-hop validity benchmark constructed for this work, the Memora role\-binding transfer setting, and Memora dense current\-state retrieval\. The recall\-preserving cascade and the protected top\-10 are defined in the v1 and v2 reports\[[1](https://arxiv.org/html/2606.26753#bib.bib1),[2](https://arxiv.org/html/2606.26753#bib.bib2)\]; v3 treats them as the fixed substrate it annotates\.

## 4The Validity Problem

#### Relevance is not validity\.

Given a query about a target entity and an accumulated memory store, a retriever ranks memories by relevance\. A memory can rank high on relevance and still be outdated: the user has since updated the fact\. The validity question is orthogonal: among the relevant memories, which ones a later source memory has superseded with respect to the target the query asks about\.

#### The relation is target\-conditioned\.

The relation between two memories is only well\-defined once the target proposition is fixed\. Two statements about the same person can stand in an update relation on one attribute \(employer\) and an unrelated relation on another \(hometown\)\. A relation judgment that ignores the target conflates these\. v3 therefore conditions every relation judgment on the specific target, which is the property that separates it from a query\-blind relation classifier \(§[11](https://arxiv.org/html/2606.26753#S11)\)\.

#### Two regimes\.

Validity reasoning takes two forms in practice\. In*role binding*, a single later memory directly rebinds a target’s attribute, and the verifier must decide whether a candidate is the superseded binding\. In*multi\-hop propagation*, an update at one node invalidates a chain of downstream memories that depended on it, and validity must propagate along the dependency structure\. v3 addresses role binding as its load\-bearing, real\-data validated capability and validates multi\-hop propagation as a mechanism with given structure, with a documented data boundary \(§[15](https://arxiv.org/html/2606.26753#S15)\)\.

## 5The Target\-Conditioned Dual\-Evidence Gate

### 5\.1Slot heads

The verifier scores whether a source memory*directly and specifically*overturns a target memory\. The central quantity is the*slot score*: a target\-conditioned proposition score that measures whether the source acts on the specific target proposition \(role binding\), rather than merely sharing a topic\. The slot score is produced by two heads that share an input form but use different backbones\.

Both heads receive the source side and the target side rendered as a text pair\. The source side is

> SOURCE\_MEMORY: \{source\_text\} QUESTION: Does SOURCE directly and specifically overturn TARGET?

paired with the target side

> TARGET\_MEMORY: \{target\_text\} DOMAIN: \{domain\}

TheMiniLM headis a binary cross\-encoder over this pair, producingmmini=P​\(direct\_overturn\)m\_\{\\mathrm\{mini\}\}=P\(\\text\{direct\\\_overturn\}\)\. TheDeBERTa\-v3 headtakes the same input form, but its backbone is initialized from an NLI checkpoint and the three\-way entailment head is replaced by a binary classification head, so it produces a comparablemnli=P​\(direct\_overturn\)m\_\{\\mathrm\{nli\}\}=P\(\\text\{direct\\\_overturn\}\)rather than raw entail/neutral/contradict logits\. The slot score is the product of the two heads:

slot=mmini⋅mnli\.\\mathrm\{slot\}\\;=\\;m\_\{\\mathrm\{mini\}\}\\cdot m\_\{\\mathrm\{nli\}\}\.\(1\)The product form is conservative: both heads must agree that the source specifically overturns the target proposition for the slot score to be high\.

### 5\.2Conservative event/operation gate

The slot score answers “does the source act on this target proposition\.” Two further factors answer “is the source an update event at all” and “is the operation an overturn\.” Letpe​v=P​\(source is an update/conflict event\)p\_\{ev\}=P\(\\text\{source is an update/conflict event\}\)andpo​p=P​\(operation type is overturn\)p\_\{op\}=P\(\\text\{operation type is overturn\}\)\. These two factors are produced by an event head and an operation head from the earlier three\-head attachment model, separate from the two slot heads of §[5\.1](https://arxiv.org/html/2606.26753#S5.SS1)\. The original formulation multiplied three independent factors,

edge=pe​v⋅slot⋅po​p,\\mathrm\{edge\}\\;=\\;p\_\{ev\}\\cdot\\mathrm\{slot\}\\cdot p\_\{op\},\(2\)with the slot factor itself the product of Equation \([1](https://arxiv.org/html/2606.26753#S5.E1)\)\. A merge analysis \(§[5\.5](https://arxiv.org/html/2606.26753#S5.SS5)\) showed that the two noisy event/operation heads need not be multiplied independently: collapsing them into a single conservative gate is statistically equivalent, and the frozen dual\-gate form is

edge=slot⋅min⁡\(pe​v,po​p\)\.\\mathrm\{edge\}\\;=\\;\\mathrm\{slot\}\\cdot\\min\(p\_\{ev\},\\,p\_\{op\}\)\.\(3\)Themin\\mingate is conservative in the same spirit as the slot product: the edge fires only when the weakest of the two evidence factors supports it\. Figure[1](https://arxiv.org/html/2606.26753#S5.F1)shows the full data flow\.

![Refer to caption](https://arxiv.org/html/2606.26753v1/x1.png)Figure 1:The target\-conditioned dual\-evidence gate\. A \(target, source\) pair, with domain, is scored by two slot heads \(MiniLM and DeBERTa\-v3\) sharing the same input form; their product is the target\-conditioned slot score \(Equation \([1](https://arxiv.org/html/2606.26753#S5.E1)\)\)\. The slot score is gated by the conservativemin⁡\(pe​v,po​p\)\\min\(p\_\{ev\},p\_\{op\}\)factor \(Equation \([3](https://arxiv.org/html/2606.26753#S5.E3)\)\) to form an edge score, and edge scores are combined over sources and over the sample by noisy\-or \(Equation \([4](https://arxiv.org/html/2606.26753#S5.E4)\)\) into an invalidity score\.
### 5\.3Aggregation by noisy\-or

A target may be overturned by any of several sources\. Edge scores from the same source are combined, and source scores are combined over the sample, by noisy\-or:

noisy​\_​or​\(\{xi\}\)=1−∏i\(1−xi\),source​\_​score=noisy​\_​or​\(edges of one source\),invalid​\_​score=noisy​\_​or​\(source scores\)\.\\mathrm\{noisy\\\_or\}\(\\\{x\_\{i\}\\\}\)\\;=\\;1\-\\prod\_\{i\}\(1\-x\_\{i\}\),\\qquad\\begin\{aligned\} \\mathrm\{source\\\_score\}&=\\mathrm\{noisy\\\_or\}\(\\text\{edges of one source\}\),\\\\ \\mathrm\{invalid\\\_score\}&=\\mathrm\{noisy\\\_or\}\(\\text\{source scores\}\)\.\\end\{aligned\}\(4\)The phrase “full product\+\+noisy\-or” denotes this composition precisely: the MiniLM and DeBERTa heads are combined by*product*\(Equation \([1](https://arxiv.org/html/2606.26753#S5.E1)\)\), and the resulting edges are combined across sources by*noisy\-or*\(Equation \([4](https://arxiv.org/html/2606.26753#S5.E4)\)\)\. The two heads are never combined by noisy\-or with each other\. A per\-seed, per\-fold decision threshold is selected on a validation split, optimizing overall accuracy jointly with invalidated\-case and valid\-case accuracy\. Algorithm[1](https://arxiv.org/html/2606.26753#alg1)states the full scoring path\.

Algorithm 1Dual\-evidence gate: target invalidity score1:target memory

tt; candidate sources

𝒮\\mathcal\{S\}; domain

dd
2:invalidity score for

tt
3:

source\_scores←\[\]\\textit\{source\\\_scores\}\\leftarrow\[\\,\]
4:foreach source

s∈𝒮s\\in\\mathcal\{S\}do

5:

x←x\\leftarrowrender pair

\(s,t,d\)\(s,t,d\)in the slot input form

6:

mmini←m\_\{\\mathrm\{mini\}\}\\leftarrowMiniLM head

\(x\)\(x\)⊳\\trianglerightP​\(direct\_overturn\)P\(\\text\{direct\\\_overturn\}\)

7:

mnli←m\_\{\\mathrm\{nli\}\}\\leftarrowDeBERTa\-v3 head

\(x\)\(x\)⊳\\trianglerightNLI backbone, binary head

8:

slot←mmini⋅mnli\\mathrm\{slot\}\\leftarrow m\_\{\\mathrm\{mini\}\}\\cdot m\_\{\\mathrm\{nli\}\}
9:

pe​v←p\_\{ev\}\\leftarrowevent head

\(x\)\(x\)⊳\\trianglerightP​\(update/conflict event\)P\(\\text\{update/conflict event\}\)

10:

po​p←p\_\{op\}\\leftarrowoperation head

\(x\)\(x\)⊳\\trianglerightP​\(operation is overturn\)P\(\\text\{operation is overturn\}\)

11:

edge←slot⋅min⁡\(pe​v,po​p\)\\mathrm\{edge\}\\leftarrow\\mathrm\{slot\}\\cdot\\min\(p\_\{ev\},p\_\{op\}\)
12:append

noisy​\_​or​\(\{edge\}\)\\mathrm\{noisy\\\_or\}\(\\\{\\mathrm\{edge\}\\\}\)tosource\_scores

13:endfor

14:return

noisy​\_​or​\(source\_scores\)\\mathrm\{noisy\\\_or\}\(\\textit\{source\\\_scores\}\)

### 5\.4A worked example

A concrete case makes the gate’s behavior explicit\. The scores below come from the graph\-line edge\-score probe, the four\-head diagnostic scorer of Equations \([1](https://arxiv.org/html/2606.26753#S5.E1)\)–\([2](https://arxiv.org/html/2606.26753#S5.E2)\), which exposes the component heads that the released single calibrator folds together \(§[6\.4](https://arxiv.org/html/2606.26753#S6.SS4)\)\. Take the source evidence*“Policy Z12 eligibility criteria revised; claims must now include additional documentation”*and the target memory*“Claim \#5241 meets eligibility criteria under Policy Z12\.”*The source is an update event acting by overturn, so the event and operation factors are high \(pe​v=0\.9968p\_\{ev\}=0\.9968,po​p=0\.9933p\_\{op\}=0\.9933\)\. Both slot heads agree that the source specifically overturns the target proposition \(mmini=0\.99995m\_\{\\mathrm\{mini\}\}=0\.99995,mnli=0\.99888m\_\{\\mathrm\{nli\}\}=0\.99888\), so the slot product is near unity \(slot≈0\.9988\\mathrm\{slot\}\\approx 0\.9988\)\. The two edge forms agree on this case: the four\-factor diagnosticpe​v⋅po​p⋅mmini⋅mnli≈0\.9890p\_\{ev\}\\cdot p\_\{op\}\\cdot m\_\{\\mathrm\{mini\}\}\\cdot m\_\{\\mathrm\{nli\}\}\\approx 0\.9890, and the frozen merged gateslot⋅min⁡\(pe​v,po​p\)≈0\.9921\\mathrm\{slot\}\\cdot\\min\(p\_\{ev\},p\_\{op\}\)\\approx 0\.9921, both place the edge near unity, matching the no\-measurable\-loss merge of §[5\.5](https://arxiv.org/html/2606.26753#S5.SS5)\. The target is surfaced aspossibly\_outdatedincontextmode and is demotable under the explicitdemoteopt\-in\. The discriminating factor is the slot product: a source that were topically related but did not act on this specific eligibility proposition would score low on the slot heads, driving the product toward zero regardless of the event factor\. This is the scope\-mismatch hard negative that stratified supervision targets \(§[6\.2](https://arxiv.org/html/2606.26753#S6.SS2)\)\.

### 5\.5Architecture freeze: head merge and the load\-bearing spine

Two architectural questions were settled by paired ablation with 95% confidence intervals on the synthetic benchmark\.

#### The event and operation heads merge with no measurable loss\.

Replacing the independent productpe​v⋅po​pp\_\{ev\}\\cdot p\_\{op\}with the conservative gatemin⁡\(pe​v,po​p\)\\min\(p\_\{ev\},p\_\{op\}\)of Equation \([3](https://arxiv.org/html/2606.26753#S5.E3)\) changes accuracy by\+0\.00%\+0\.00\\%with 95% CI\[−0\.88%,\+0\.84%\]\[\-0\.88\\%,\+0\.84\\%\]\. The merged gate is the frozen dual\-gate architecture: simpler and statistically indistinguishable from the two\-head form\.

#### The slot product is the spine\.

Scoring with the slot\-only product and noisy\-or gives−1\.00%\-1\.00\\%relative to the full configuration \(95% CI\[−2\.19%,\+0\.20%\]\[\-2\.19\\%,\+0\.20\\%\]\), within noise: the slot product carries essentially all of the signal\. Removing the slot mechanism entirely \(event/operation heads without the slot interaction\) costs about 14 percentage points\. The target\-conditioned slot product is therefore the load\-bearing spine, and the event/operation heads are mergeable, conservative overhead\. Table[1](https://arxiv.org/html/2606.26753#S5.T1)collects the ablation\.

Table 1:Architecture ablation on the synthetic multi\-hop benchmark \(paired, 95% CI\)\. The merged event/operation gate matches the full two\-head model; the slot\-only product matches the full configuration within noise; removing the slot interaction costs about 14 points\. The merged\-gate, slot\-spine architecture is the frozen research\-verifier configuration\.

## 6Training the Verifier

### 6\.1Objective

Each slot head is trained as a binary classifier on labeled \(source, target, domain\) pairs\. A two\-logit head uses weighted cross\-entropy; a single\-logit head uses binary cross\-entropy with logits:

ℒslot=CEw​\(y^,y\),y∈\{overturn,non\-overturn\}\.\\mathcal\{L\}\_\{\\mathrm\{slot\}\}\\;=\\;\\mathrm\{CE\}\_\{w\}\\big\(\\hat\{y\},\\,y\\big\),\\qquad y\\in\\\{\\text\{overturn\},\\ \\text\{non\-overturn\}\\\}\.\(5\)The class weighting addresses the imbalance between the relatively rare direct overturns and the abundant non\-overturns\.

### 6\.2Stratified supervision

The discriminating signal lives in the hard cases, so the training distribution is stratified rather than sampled uniformly:

- •Clean positive: a GPT\-confirmed direct text overturn\.
- •Hard negative: scope mismatch or temporal non\-conflict, namely a source that looks superficially like an overturn but does not act on the target proposition\.
- •Easy negative: an ordinary unrelated pair\.

Clean positives are confirmed by a GPT labeler \(gpt\-4\.1\) that judges whether the source directly overturns the target\. These labels supervise training only; no GPT label enters the evaluation sets, whose targets come from the synthetic\-benchmark construction and the Memora ground truth\. Hard negatives are oversampled by a focus multiplier \(defaulthard\_focus\_multiplier=4\\textit\{hard\\\_focus\\\_multiplier\}=4\), and low\-overlap or failure\-subtype positives are oversampled \(defaultpos\_multiplier=3\\textit\{pos\\\_multiplier\}=3\)\. Stratifying the signal so that rule\-easy and rule\-hard cases contribute distinct gradients, rather than being averaged, produced the first large jump to84\.5%±2\.884\.5\\%\\pm 2\.8\(§[8](https://arxiv.org/html/2606.26753#S8)\)\.

### 6\.3The real\-data feedback loop

The transfer capability comes from a feedback loop that analyzes real failure patterns, generates synthetic supervision matching them, and transfers back with zero target\-side labels\. Algorithm[2](https://arxiv.org/html/2606.26753#alg2)states the loop\.

Algorithm 2Real\-data feedback loop \(zero target\-side labels at transfer\)1:Analyze Memora role\-binding failures of a target\-position rule and zero\-shot NLI to identify recurring hard\-case patterns

2:Generate synthetic paired supervision \(clean positive vs scope/temporal hard negative\) matching those patterns, without Memora labels

3:Refine the slot heads on the synthetic pairs under the stratified objective \(§[6\.2](https://arxiv.org/html/2606.26753#S6.SS2)\)

4:Evaluate transfer on Memora with no target\-side training labels

5:returnrefined verifier

This loop is the strongest validated capability in v3: a verifier trained without Memora target labels transfers to the full Memora role\-binding set at98\.8%±0\.998\.8\\%\\pm 0\.9group\-all\-correct \(§[9](https://arxiv.org/html/2606.26753#S9)\)\.

#### What zero target\-side labels means\.

The transfer is label\-free in a precise sense: training uses synthetic pairs only, and no Memora ground\-truth label enters the objective\. Exact group, row, and source\-text overlap between the synthetic training pool and the Memora evaluation set is zero\. Two couplings remain at the level of surface form rather than supervision, and we state them directly\. First, five entity names \(a small set of public\-figure and topic strings\) appear in both pools as surface strings; the Memora evaluation instances and their labels stay outside training\. Second, the synthetic hard\-negative templates \(§[6\.2](https://arxiv.org/html/2606.26753#S6.SS2)\) are designed from an error analysis of Memora role\-binding failures, which transfers the*shape*of the hard cases without transferring their labels\. The transfer is therefore label\-free, and the role\-binding numbers hold under this exact sense\.

### 6\.4The deployed query\-conditioned calibrator

The releaseddemotemode does not deploy the synthetic\-graph scorer\. It uses a query/source/target cross\-encoder calibrator whose input is conditioned on the user query:

> USER\_QUERY: \{query\} SOURCE\_EVIDENCE: \{source\} TASK: Decide whether the target memory should be demoted for this user query\.

paired withTARGET\_MEMORY: \{target\}\. It is trained with a weighted cross\-entropy plus an event\-role margin term,

ℒ=CEw​\(y^,y\)\+λrole⋅ℒrole​\_​margin,\\mathcal\{L\}\\;=\\;\\mathrm\{CE\}\_\{w\}\\big\(\\hat\{y\},y\\big\)\\;\+\\;\\lambda\_\{\\mathrm\{role\}\}\\cdot\\mathcal\{L\}\_\{\\mathrm\{role\\\_margin\}\},\(6\)whereℒrole​\_​margin\\mathcal\{L\}\_\{\\mathrm\{role\\\_margin\}\}requires that, within one event group, the rows that should be demoted \(old or current bindings under update\) score above the rows that must be protected\. The margin term is what makes the calibrator query\-conditioned rather than query\-blind, which is the property that the never\-demote and query\-blind baselines lack \(§[11](https://arxiv.org/html/2606.26753#S11)\)\.

## 7Experimental Protocol

#### Settings\.

Real\-distribution validity behavior is evaluated on Memora\[[15](https://arxiv.org/html/2606.26753#bib.bib15)\], a public long\-term memory benchmark introduced by Uddin et al\., spanning weeks\-to\-months user conversations with explicit memory mutation and a Forgetting\-Aware Memory Accuracy \(FAMA\) metric that penalizes reliance on obsolete memory\. The A1 role\-binding and dense\-retrieval demotion slices used here are derived from the released Memora traces\. Memora labels are not used as training targets; Memora residual errors are used only to identify failure patterns for synthetic template design \(§[6\.3](https://arxiv.org/html/2606.26753#S6.SS3)\)\. Three evaluation settings are used\. The*synthetic multi\-hop validity benchmark*constructs invalidation chains of controlled length, where an update at one node should invalidate downstream nodes that depend on it; it is the setting in which the dual\-evidence gate and the propagation mechanism are measured with known structure\. The*Memora A1 role\-binding transfer*setting evaluates whether a candidate is the superseded binding of a target whose attribute a later source rebinds; it is the real\-data setting and uses zero target\-side training labels\. The*Memora dense current\-state retrieval*setting asks for a target’s current attribute against a store containing both the current binding and superseded ones; it is the setting in which the deployed validity layer is measured end\-to\-end\.

#### Seeds and aggregation\.

Method\-level tables use seeds\[7,11,23,31,47\]\[7,11,23,31,47\]and report mean±\\pmstandard deviation\. The dense\-retrieval method\-level evaluation scores69,20069\{,\}200source\-query rows across the five seeds, with34,60034\{,\}600top\-1 summary rows\. The released\-checkpoint benchmark is a single checkpoint over6,9206\{,\}920source\-query rows and20,76020\{,\}760target predictions: each source\-query row is scored against the candidate targets in its group, so one row expands to several\(query, source, target\)scoring instances \(here20,760/6,920=320\{,\}760/6\{,\}920=3on average\)\. The Memora A1 role\-binding set has 173 groups, with a 37\-group rule\-hard subset and a 12\-group both\-failed subset; the wider intervals on these subsets reflect their size\.

#### Metrics\.

On role binding the metric is group\-all\-correct accuracy: a group counts as correct only when every binding decision in the group is correct\. On dense retrieval the metrics are pair accuracy \(target\-conditioned relation accuracy\), demote recall \(fraction of superseded memories correctly demoted\), protect recall \(fraction of non\-superseded memories correctly preserved\), and current\-active Hit@1 \(H@1, end\-task retrieval of the current binding at rank 1\)\. On the synthetic benchmark the metric is overall accuracy, reported jointly with invalidated\-case and valid\-case accuracy so that aggressive propagation cannot trade valid\-case correctness for invalidated\-case recall undetected\.

#### Threshold selection\.

The dual\-evidence gate selects a decision threshold per seed and per fold on a validation split, optimizing overall accuracy jointly with invalidated\-case and valid\-case accuracy \(§[5\.3](https://arxiv.org/html/2606.26753#S5.SS3)\)\. Thresholds are never selected on the test split\.

#### Isolation\.

Evaluation\-only fields \(gold labels, staleness flags, teacher scores, GPT labels, and entity/slot identifiers\) are used only as training or evaluation targets and never as inference features; the deployed API rejects them at inference \(§[14\.2](https://arxiv.org/html/2606.26753#S14.SS2)\)\. The synthetic\-graph scorer and the deployed query\-conditioned calibrator are trained separately, and the calibrator uses no synthetic\-graph labels at inference\.

## 8Verifier Progression

Figure[2](https://arxiv.org/html/2606.26753#S8.F2)traces the verifier across three milestones on the synthetic benchmark\. The first end\-to\-end system that turned text into a dependency graph and propagated over it reached an overall65\.2%65\.2\\%, with visible edge\-error amplification along chains\. Clean real\-slot supervision \(§[6\.2](https://arxiv.org/html/2606.26753#S6.SS2)\) lifted accuracy to84\.5%±2\.884\.5\\%\\pm 2\.8\. The corrected dual\-evidence gate, which fixed a duplicate\-edge deduplication bug and applies Equations \([1](https://arxiv.org/html/2606.26753#S5.E1)\)–\([4](https://arxiv.org/html/2606.26753#S5.E4)\), reached90\.12%±1\.7390\.12\\%\\pm 1\.73, the synthetic headline\. A nearby run measured90\.20%±2\.1990\.20\\%\\pm 2\.19before this correction; the corrected90\.12%±1\.7390\.12\\%\\pm 1\.73is the figure used throughout\.

![Refer to caption](https://arxiv.org/html/2606.26753v1/x2.png)Figure 2:Verifier progression\. \(a\) Synthetic multi\-hop accuracy across milestones: the65\.2%65\.2\\%end\-to\-end learned\-graph baseline, the84\.5%±2\.884\.5\\%\\pm 2\.8clean\-supervision milestone, and the90\.12%±1\.7390\.12\\%\\pm 1\.73corrected dual\-evidence gate\. \(b\) Memora role\-binding transfer \(full set, zero target\-side labels\) from the first transfer checkpoint \(90\.4%90\.4\\%\) to the final DeBERTa verifier \(98\.8%±0\.998\.8\\%\\pm 0\.9\), showing how the synthetic\-side mechanism converts into real\-data transfer\.
## 9Role\-Binding Verification

The strongest validated capability is target\-conditioned role binding: deciding, for a target whose attribute a later source rebinds, whether a candidate memory is the superseded binding\. Table[2](https://arxiv.org/html/2606.26753#S9.T2)reports group\-all\-correct accuracy on the Memora A1 transfer set \(173 groups\), the rule\-hard subset \(37 groups where a target\-position rule fails by construction\), and the both\-failed subset \(12 groups where the position rule and zero\-shot NLI both fail\)\. All trained rows use zero Memora target\-side labels\.

Table 2:Target\-conditioned role binding on Memora A1 transfer \(group\-all\-correct accuracy\)\. all\-173 is the full set; rule\-hard\-37 is the subset where a target\-position rule fails by construction; both\-failed\-12 is the subset where the position rule and zero\-shot NLI both fail\. v3 rows use zero Memora target\-side training labels; final rows are 5\-seed mean±\\pmstd\. A dash \(—\) marks a split that is undefined for that baseline\.The verifier holds up exactly where the cheap signals collapse\. A target\-position rule scores78\.6%78\.6\\%overall but0%0\\%on the rule\-hard subset, which is hard by construction; zero\-shot NLI scores64\.2%64\.2\\%overall and0%0\\%on the both\-failed subset\. The relevance cross\-encoder scores17\.9%17\.9\\%, confirming that relevance ranking and validity verification are distinct tasks\. The first transfer checkpoint already reached90\.4/81\.1/66\.790\.4/81\.1/66\.7across the three splits; the final DeBERTa verifier with merged supervision reaches98\.8/96\.2/93\.398\.8/96\.2/93\.3, with the wider intervals on the small hard subsets reflecting their size \(37 and 12 groups\)\.

## 10Multi\-Hop Validity Propagation

When an update invalidates a chain of dependent memories, validity must propagate along the dependency structure\. Figure[3](https://arxiv.org/html/2606.26753#S10.F3)and Table[3](https://arxiv.org/html/2606.26753#S10.T3)report a diagnostic on synthetic invalidated chains of length 2, 3, and 4: the fraction of the chain correctly marked invalid by single\-hop scoring, by propagation over a learned graph, and by an oracle that propagates over the true structure\.

![Refer to caption](https://arxiv.org/html/2606.26753v1/x3.png)Figure 3:Multi\-hop validity propagation on synthetic invalidated chains \(invalidated\-only\)\. Single\-hop scoring captures a small, roughly flat fraction of the chain; propagation over a learned graph recovers most of it; oracle propagation over the true structure is the ceiling\.Table 3:Multi\-hop validity propagation\. The upper block reports all chains by length \(recency, v2\-relevance, single\-text, single\-hop, learned\-graph, oracle\); the lower block reports the invalidated\-only fraction correctly invalidated\. Learned\-graph rows are 5\-seed mean±\\pmstd\. Per\-chain sample sizes are: all chainsn=159/170/173n=159/170/173at lengths2/3/42/3/4; invalidated\-onlyn=110/75/62n=110/75/62; valid controlsn=49/95/111n=49/95/111\.On the invalidated\-only split, single\-hop scoring captures only a small fraction of the chain \(6–11%\) regardless of length, because a single\-hop judgment cannot see that an upstream update has invalidated a node two or three hops away\. Propagation over a learned graph recovers most of the chain \(68–80%\), demonstrating that propagation is the right structure for multi\-hop validity\. The two blocks trend in opposite directions with chain length, and the reason is the shifting composition of each split: as chains lengthen, the all\-chains set is increasingly dominated by valid controls \(the valid\-control count rises from 49 to 111 across lengths22to44\), which pulls the all\-chains learned\-graph figure down, while the invalidated\-only set shrinks ton=62n=62at length 4\. The invalidated\-only rebound to80\.3%80\.3\\%at length 4 should not be over\-read: it rests on the smallest per\-chain sample and carries the widest interval \(±5\.7\\pm 5\.7\), consistent with small\-sample variation rather than a genuine increase in propagation accuracy with depth\. The gap to the oracle, together with false\-overturn harm on the valid controls, scopes propagation to the given\-structure regime: with a reliable dependency structure, propagation works; constructing that structure fully automatically is the boundary characterized in §[15](https://arxiv.org/html/2606.26753#S15)\.

## 11The Validity Context Layer in Dense Retrieval

The deployed layer is evaluated on Memora dense current\-state retrieval, where a query asks for a target’s current attribute and the store contains both the current binding and superseded ones\. Table[4](https://arxiv.org/html/2606.26753#S11.T4)reports method\-level results \(5 seeds,69,20069\{,\}200source\-query rows scored across seeds\), with the released\-checkpoint figures reported alongside in §[14\.4](https://arxiv.org/html/2606.26753#S14.SS4)\.

Table 4:Validity context layer on Memora dense current\-state retrieval \(5\-seed mean±\\pmstd\)\. never\-demote is the no\-validity baseline; query\-blind is a relation classifier that ignores the target and matches never\-demote exactly; top1\-retrieved\-source is the deployed policy; oracle\-source uses the true superseding source\. pair acc is target\-conditioned relation accuracy; demote/protect recall measure superseded vs non\-superseded handling; current\-active H@1 is end\-task retrieval of the current binding\.The never\-demote baseline retrieves the current binding only45\.1%45\.1\\%of the time, because superseded memories outrank the current one on relevance alone\. A query\-blind relation classifier, one that judges the relation without conditioning on the target, matches the never\-demote numbers exactly: knowing the relation without the query intent collapses to protecting every memory \(0%0\\%demote recall,100%100\\%protect recall\), so it never recovers the current binding\. This is the empirical case for target conditioning: the relation signal is only useful once it is tied to the target the query asks about\. The deployed v3 layer, using the top\-1 retrieved source as the superseding evidence, raises current\-active H@1 to95\.7%±1\.295\.7\\%\\pm 1\.2while protecting non\-superseded memories at99\.4%99\.4\\%recall and demoting superseded ones at92\.9%92\.9\\%recall\. The oracle source \(using the true superseding memory\) reaches100%100\\%H@1, locating the remaining headroom in source selection rather than in the relation judgment itself\.

#### Old\-target and event splits\.

On the same setting the layer reaches92\.8%±1\.192\.8\\%\\pm 1\.1on the old\-target all\-type split and89\.1%±1\.389\.1\\%\\pm 1\.3on the event all\-type split, consistent with role binding being the strongest regime and event\-style updates the harder one\.

## 12LoCoMo Density Audit

ConvMemory v1 and v2 are evaluated on LoCoMo\[[16](https://arxiv.org/html/2606.26753#bib.bib16)\], so a natural question is how the validity layer behaves there\. The answer is a finding about the data rather than a benchmark result: on LoCoMo, supersession is rare enough that the layer’s correct default is to annotate, not to demote\. A high\-value GPT audit of 100 candidate cases found valid demotions in33of them \(3\.0%3\.0\\%yield\), and2020negative controls produced0false positives\. The audited positive set is sparse \(on the order of a handful of confidently superseded memories across the audit\), which is too thin to support a method\-level demotion benchmark\. The audit is used to decide deployment scope, not to estimate method\-level accuracy\. Accordingly, thecontextdefault preserves ranking on LoCoMo with no behavior change, and the opt\-indemotemode shows no reliable gain there\. LoCoMo is therefore evidence that real conversational supersession density can be low \(exactly the regime in which annotate\-by\-default is the safe policy\), rather than a validation set for demotion\. The settings where demotion is validated \(Memora dense current\-state retrieval\) are those with enough supersession density to measure it\.

## 13Cost\-Conscious Routing

This section analyzes the research dual\-gate verifier, not the released single\-checkpoint calibrator of §[14\.4](https://arxiv.org/html/2606.26753#S14.SS4)\. The dual\-evidence gate runs both slot heads on every pair\. A cost\-conscious router \(B1\) instead runs the cheap MiniLM head first and routes to the DeBERTa head only when the cheap head is uncertain\. On the synthetic benchmark, the best router \(a fast\-accept variant with agreement gating, combined by noisy\-or\) reaches89\.56%±2\.5489\.56\\%\\pm 2\.54, indistinguishable within seed\-level variance from the full\-product anchor in the same run \(89\.40%±2\.7989\.40\\%\\pm 2\.79\) and from the corrected gate \(90\.12%±1\.7390\.12\\%\\pm 1\.73\)\.

The cost behavior, however, depends on the data, not the router\. On the synthetic benchmark the router still sends99\.5%99\.5\\%of pairs to the DeBERTa head, because synthetic data has almost no easy mass by construction\. On real Memora data the router preserves the transfer numbers \(90\.4/81\.1/66\.790\.4/81\.1/66\.7\) while sending far fewer pairs to the expensive head:75\.0%75\.0\\%at the tightest agreement threshold,86\.0%86\.0\\%at a looser one, and94\.5%94\.5\\%at the loosest\. The reframing follows directly:*cost saving depends on the easy mass in the deployment distribution, not on the router alone*\. A router added to a distribution with no easy mass saves nothing; the same router on real data, which contains easy mass, routes a meaningful fraction away from the expensive head\. Quantified against the component latencies of Table[8](https://arxiv.org/html/2606.26753#S14.T8), routing on real data reduces the full dual\-gate cost from about18\.918\.9to about15\.115\.1ms per top\-10 query, while on synthetic data the near\-total routing leaves it essentially unchanged at about18\.818\.8ms\.

## 14Deployment: Modes, Contracts, and Cost

### 14\.1Modes

The layer exposes four accepted values through a singlevalidity\_modeargument, corresponding to three behavioral modes \(Noneandoffshare one behavior\)\.

- •None/off: ConvMemory behavior is unchanged; the v1/v2 path is byte\-identical to a build without the v3 module\.
- •context\(default\): the layer attaches structured validity metadata to each result while preserving the candidate set and the rank order\. The agent reads the validity context note and thepossibly\_outdatedflag and decides how to use them\.
- •demote: an explicit opt\-in that reorders results for dense current\-state workloads, preserving the candidate set and result count while moving superseded memories down\. This is the only mode in which v3 changes ordering\.

### 14\.2Safety contracts

Six properties are enforced by named tests rather than asserted in prose\. The full suite reaches 41 passing tests\. Table[5](https://arxiv.org/html/2606.26753#S14.T5)maps each contract to its enforcing test\.

Table 5:Validity\-layer safety contracts, each enforced by a named test\. The full suite reaches 41 passing tests; the supplementary tests below the rule cover batching, format, and round\-trip behavior\.The input contract rejects a fixed set of evaluation\-only fields\. If any candidate passed at inference contains any of the following keys, the API raisesValueError:

> gold,gold\_ids,is\_current,is\_latest,is\_stale,stale,answer,answer\_text,ce\_score,mxbai\_score,teacher\_score,gpt\_label,entity\_id,slot\_id\.

The output contract sanitizes context evidence so that none of these fields leaks into the validity metadata returned to the agent\.

### 14\.3Cost and configuration

The released default checkpoint is a DeBERTa\-v3 NLI backbone with184,423,682184\{,\}423\{,\}682parameters\. The default package mode isvalidity\_mode="context"; demotion is an explicit opt\-in\. The source policy is top\-1\. As a released package measurement, the batched public API at batch size 512 scores6,9206\{,\}920source\-query rows at1\.58441\.5844ms per source\-query pair, with a module load time of2\.162\.16s\. An earlier unbatched smoke path through the same module measured17\.58717\.587ms per pair and is retained only as a non\-headline sanity check; the batched figure is the headline cost\. Annotating a top\-10 context costs on the order of10×1\.584410\\times 1\.5844ms of scorer time before retrieval and source\-selection overhead; the measured unit is per pair\.

Table 6:Released package configuration and measured cost \(public API, batch size 512\)\. The public API unit is one source\-query row; internal scoring expands to\(query, source, target\)instances\.#### Component latency and configuration estimates\.

A separate component latency probe measures each slot head in isolation across batch sizes, which supports estimating the dual\-gate and routed\-gate costs from measured components\. This probe is distinct from the released package measurement above: it scores a fixed probe set per batch size rather than running the full public API benchmark, so its per\-pair figures are component timings rather than a single deployed cost\. Table[7](https://arxiv.org/html/2606.26753#S14.T7)reports the component timings, and Table[8](https://arxiv.org/html/2606.26753#S14.T8)reports the configuration estimates built from them\.

Table 7:Component scoring latency from the component probe\. Each slot head is timed in isolation; “best batched” is the fastest batch size found for that head in this probe\.Table 8:Configuration cost estimated from the component latencies of Table[7](https://arxiv.org/html/2606.26753#S14.T7)\(batch=32=32\)\. The dual\-gate and routed\-gate rows are estimated asdual=MiniLM\+DeBERTa\\mathrm\{dual\}=\\mathrm\{MiniLM\}\+\\mathrm\{DeBERTa\}androuted=MiniLM\+r⋅DeBERTa\\mathrm\{routed\}=\\mathrm\{MiniLM\}\+r\\cdot\\mathrm\{DeBERTa\}for routing fractionrr, not measured as direct end\-to\-end wall\-clock\.The dual\-gate and routed\-gate figures are estimated from the measured MiniLM and DeBERTa component latencies rather than measured as a single end\-to\-end wall\-clock run\. On real data, where a quarter of pairs are routed away from the expensive head, the routed gate is estimated at15\.1315\.13ms per top\-10 query against the full dual gate’s18\.8818\.88; on synthetic data, where nearly all pairs reach the expensive head, the routed gate stays at18\.8018\.80, essentially unchanged\. This quantifies the easy\-mass effect of §[13](https://arxiv.org/html/2606.26753#S13): the routing saving tracks the easy mass of the deployment distribution\.

### 14\.4Released checkpoint figures

The released checkpoint reproduces the method\-level dense\-retrieval behavior\. On the public API/package benchmark \(single checkpoint,6,9206\{,\}920source\-query rows,20,76020\{,\}760target predictions\) it reaches pair accuracy98\.7%98\.7\\%, demote recall93\.6%93\.6\\%, protect recall99\.4%99\.4\\%, old\-target all\-type93\.1%93\.1\\%, event all\-type89\.6%89\.6\\%, and current\-active H@196\.5%96\.5\\%, at1\.58441\.5844ms per source\-query pair\. These are the figures a user reproduces from the published checkpoint; the figures in Table[4](https://arxiv.org/html/2606.26753#S11.T4)are the 5\-seed method\-level estimates\. The provenance relationship is detailed in Appendix[B](https://arxiv.org/html/2606.26753#A2)\.

## 15Boundary Findings

Two boundaries are reported as findings, each scoping a capability precisely\.

#### Multi\-hop relation labels for free\-form conversation are scarce\.

The verifier is validated on role binding with real transfer \(Memora\) and on multi\-hop propagation with synthetic structure\. Target\-conditioned relation labels for free\-form multi\-hop conversation \(the data needed to validate end\-to\-end multi\-hop propagation on natural dialogue\) are not yet available at the scale required, to our knowledge\. v3 therefore validates multi\-hop propagation in the given\-structure regime and reports natural\-dialogue multi\-hop validation as the next data milestone\.

#### Strict prerequisite edges require counterfactual necessity\.

An extensive line of experiments on automatic dependency\-graph construction concluded that strict prerequisite edges \(edges asserting that one memory is necessary for another\) require a counterfactual necessity judgment grounded in world knowledge\. A discriminative relation model provides support edges reliably but does not provide strict necessity reliably, and aggressive automatic propagation introduces false\-overturn harm on valid controls \(§[10](https://arxiv.org/html/2606.26753#S10)\)\. Automatic strict\-prerequisite graph construction is therefore not part of the released default path; it is reported as a boundary and future research direction, and the deployed layer uses given or support structure rather than inferred strict prerequisites\.

#### Demotion is scoped, not default\.

Query\-conditioned demotion is validated on Memora dense current\-state retrieval, where a query asks for a target’s current attribute\. On general conversational retrieval the safe default iscontext: surface the validity evidence and let the agent decide\. Demotion is an explicit opt\-in for the dense current\-state regime where it is validated\.

## 16Reproducibility

The package is installable viapip install convmemory==0\.6\.0\(PyPI\)\. The released validity checkpoint is on the Hugging Face Hub atPurdy0228/ConvMemory\-v3\-Validity\-Context\. The build produces a clean 46 KiB wheel333The wheel is47,50447\{,\}504bytes\.with no results, checkpoints, or weights mixed in, under the MIT license\. A minimal usage example:

```
from convmemory import ConvMemory
model = ConvMemory.from_pretrained("Purdy0228/ConvMemory-LoCoMo-MPNet")
model.load_validity_module("Purdy0228/ConvMemory-v3-Validity-Context")
# context mode (default): preserves ranking, attaches validity metadata
results = model.retrieve(
    query=q,
    memories=ms,
    top_k=10,
    validity_mode="context",
)
for r in results:
    if r.validity and r.validity.get("status") == "possibly_outdated":
        print(r.validity["context_note"])
# validity_mode="demote" is an explicit opt-in that reorders results
```

## 17Discussion

ConvMemory v3 contributes a validity context layer: target\-conditioned relation verification, conservative evidence surfacing, and opt\-in query\-conditioned demotion\. The verifier transfers from synthetic structure to real role binding with zero target\-side labels, the deployed layer preserves retrieval by default and recovers current\-state accuracy under explicit opt\-in, and its behavior is pinned by machine\-verifiable contracts\. Multi\-hop graph propagation is validated as a mechanism; full automatic strict dependency\-graph construction remains a boundary, because strict necessity requires counterfactual world knowledge\.

#### Why target conditioning is the load\-bearing idea\.

Three independent results point at the same conclusion\. The architecture ablation \(§[5\.5](https://arxiv.org/html/2606.26753#S5.SS5)\) shows that removing the slot interaction costs about 14 points while merging the event and operation heads costs nothing, so the target\-conditioned slot product, not the event/operation machinery, carries the signal\. The dense\-retrieval comparison \(§[11](https://arxiv.org/html/2606.26753#S11)\) shows that a query\-blind relation classifier collapses to the never\-demote failure shape, so the relation signal is only useful once tied to the target\. The role\-binding comparison \(§[9](https://arxiv.org/html/2606.26753#S9)\) shows that a relevance cross\-encoder scores17\.9%17\.9\\%, so relevance and validity are genuinely different tasks\. Across the three settings, the property that separates a working verifier from a failing one is the same: conditioning the relation judgment on the specific target proposition\.

#### Where the headroom is\.

In dense retrieval the oracle source reaches100%100\\%current\-active H@1 against the deployed layer’s95\.7%95\.7\\%, which places the remaining headroom in source selection rather than in the relation judgment\. The relation judgment itself is near\-saturated \(pair accuracy98\.6%98\.6\\%\); the path to the oracle runs through retrieving the correct superseding source, not through a stronger relation model\.

### 17\.1Limitations

#### Multi\-hop validation uses synthetic structure\.

The propagation mechanism is validated on synthetic invalidation chains with known structure\. End\-to\-end validation on natural multi\-hop dialogue awaits target\-conditioned relation labels at scale, which are, to our knowledge, scarce \(§[15](https://arxiv.org/html/2606.26753#S15)\)\.

#### Strict\-edge construction is a boundary, not a released feature\.

The deployed layer uses given or support structure\. Fully automatic construction of strict prerequisite edges is not part of the released default path and is reported as a boundary and future research direction, because strict necessity requires counterfactual world knowledge that a discriminative model does not provide reliably\.

#### Demotion is scoped to dense current\-state retrieval\.

Thedemotemode is validated on Memora dense current\-state retrieval and is an explicit opt\-in there\. On general conversational retrieval the validated default iscontext, which surfaces evidence without reordering\.

#### Cost savings are distribution\-dependent\.

The cost\-conscious router \(§[13](https://arxiv.org/html/2606.26753#S13)\) saves compute in proportion to the easy mass of the deployment distribution\. On distributions with little easy mass, both heads run on most pairs and the saving is small; the saving is a property of the data, reported as such\.

#### Released checkpoint versus method\-level figures\.

The published checkpoint is a single representative checkpoint; the headline dense\-retrieval figures are 5\-seed method\-level estimates \(Appendix[B](https://arxiv.org/html/2606.26753#A2)\)\. A user evaluating the single checkpoint reproduces the package\-benchmark figures, which are consistent with, and measured separately from, the 5\-seed estimates\.

### 17\.2Future Work

Three directions follow\. First, target\-conditioned relation labels on natural multi\-hop dialogue would lift multi\-hop propagation from the given\-structure regime to end\-to\-end validation\. Second, source selection is the located headroom in dense retrieval: a stronger source\-retrieval policy converts directly into end\-task accuracy, since the relation judgment is already near\-saturated\. Third, the validity layer composes with an agent’s memory\-management loop, where thecontextannotations become inputs to write, read, and consolidate decisions\.

## Appendix ASource\-of\-Truth Checklist

Every quantitative result in this report traces to an internal measurement record\. Table[9](https://arxiv.org/html/2606.26753#A1.T9)maps result types to the measurement that establishes them\. The released package and checkpoint are public \(PyPI and the Hugging Face Hub\); the underlying measurement records are local artifacts retained by the author and available on request\.

Table 9:Mapping from result type to the measurement that establishes it\.
## Appendix BReleased Checkpoint Provenance

The relationship between the published checkpoint and the headline numbers follows the same three\-level convention as ConvMemory v2\.

- •Method\-level: 5\-seed estimates over seeds\[7,11,23,31,47\]\[7,11,23,31,47\],69,20069\{,\}200source\-query rows scored across seeds, with top\-1 summary rows numbering34,60034\{,\}600\. These are the headline dense\-retrieval figures \(Table[4](https://arxiv.org/html/2606.26753#S11.T4)\)\.
- •Representative checkpoint: the seed\-7 representative checkpoint exported for release\.
- •Package benchmark: the public API/package benchmark for that checkpoint,6,9206\{,\}920source\-query rows and20,76020\{,\}760target predictions \(§[14\.4](https://arxiv.org/html/2606.26753#S14.SS4)\)\.

The released checkpoint implements the method; the method\-level figures are 5\-seed estimates, and the package\-benchmark figures are the single\-checkpoint measurements a user reproduces\. The recorded artifact hashes are

> model:446ee0cf6df4a8967e1a78c46d2ff3a2d777de65efbf475d2278d99468faa8d9 config:81eddb5f2ff4545dcf4b7655fedd1f7cf846248ad8962394195e6960a2e07849

## References

- \[1\]Taiheng Pan\.ConvMemory: A Lightweight Learned Memory Reranker, a Negative Attribution Result, and a Research\-Preview Conflict Editor\.arXiv preprint arXiv:2605\.28062, 2026\.[https://arxiv\.org/abs/2605\.28062](https://arxiv.org/abs/2605.28062)\.
- \[2\]Taiheng Pan\.ConvMemory v2: A Recall\-Preserving Top\-10 Evidence Reranker for Conversational Memory Retrieval\.arXiv preprint arXiv:2606\.10842, 2026\.[https://arxiv\.org/abs/2606\.10842](https://arxiv.org/abs/2606.10842)\.
- \[3\]Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie\-Yan Liu\.MPNet: Masked and permuted pre\-training for language understanding\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, 2020\.[https://arxiv\.org/abs/2004\.09297](https://arxiv.org/abs/2004.09297)\.
- \[4\]Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou\.MiniLM: Deep self\-attention distillation for task\-agnostic compression of pre\-trained transformers\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, 2020\.[https://arxiv\.org/abs/2002\.10957](https://arxiv.org/abs/2002.10957)\.
- \[5\]Pengcheng He, Jianfeng Gao, and Weizhu Chen\.DeBERTaV3: Improving DeBERTa using ELECTRA\-style pre\-training with gradient\-disentangled embedding sharing\.arXiv preprint arXiv:2111\.09543, 2021\.[https://arxiv\.org/abs/2111\.09543](https://arxiv.org/abs/2111.09543)\.
- \[6\]Nils Reimers and Iryna Gurevych\.Sentence\-BERT: Sentence embeddings using Siamese BERT\-networks\.In*Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)*, 2019\.[https://arxiv\.org/abs/1908\.10084](https://arxiv.org/abs/1908.10084)\.
- \[7\]Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng\.MS MARCO: A human generated machine reading comprehension dataset\.arXiv preprint arXiv:1611\.09268, 2016\.[https://arxiv\.org/abs/1611\.09268](https://arxiv.org/abs/1611.09268)\.
- \[8\]James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal\.FEVER: a large\-scale dataset for fact extraction and VERification\.In*Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics \(NAACL\-HLT\)*, 2018\.[https://arxiv\.org/abs/1803\.05355](https://arxiv.org/abs/1803.05355)\.
- \[9\]Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov\.Locating and editing factual associations in GPT\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, 2022\.[https://arxiv\.org/abs/2202\.05262](https://arxiv.org/abs/2202.05262)\.
- \[10\]Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau\.Mass\-editing memory in a transformer\.arXiv preprint arXiv:2210\.07229, 2022\.[https://arxiv\.org/abs/2210\.07229](https://arxiv.org/abs/2210.07229)\.
- \[11\]Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G\. Patil, Ion Stoica, and Joseph E\. Gonzalez\.MemGPT: Towards LLMs as operating systems\.arXiv preprint arXiv:2310\.08560, 2023\.[https://arxiv\.org/abs/2310\.08560](https://arxiv.org/abs/2310.08560)\.
- \[12\]Joon Sung Park, Joseph C\. O’Brien, Carrie J\. Cai, Meredith Ringel Morris, Percy Liang, and Michael S\. Bernstein\.Generative agents: Interactive simulacra of human behavior\.In*Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology \(UIST\)*, 2023\.[https://arxiv\.org/abs/2304\.03442](https://arxiv.org/abs/2304.03442)\.
- \[13\]Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang\.MemoryBank: Enhancing large language models with long\-term memory\.In*Proceedings of the AAAI Conference on Artificial Intelligence*, 2024\.[https://arxiv\.org/abs/2305\.10250](https://arxiv.org/abs/2305.10250)\.
- \[14\]Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav\.Mem0: Building production\-ready AI agents with scalable long\-term memory\.arXiv preprint arXiv:2504\.19413, 2025\.[https://arxiv\.org/abs/2504\.19413](https://arxiv.org/abs/2504.19413)\.
- \[15\]Md Nayem Uddin, Kumar Shubham, Eduardo Blanco, Chitta Baral, and Gengyu Wang\.From recall to forgetting: Benchmarking long\-term memory for personalized agents\.arXiv preprint arXiv:2604\.20006, 2026\.[https://arxiv\.org/abs/2604\.20006](https://arxiv.org/abs/2604.20006)\.
- \[16\]Adyasha Maharana, Dong\-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang\.Evaluating very long\-term conversational memory of LLM agents\.arXiv preprint arXiv:2402\.17753, 2024\.[https://arxiv\.org/abs/2402\.17753](https://arxiv.org/abs/2402.17753)\.
- \[17\]Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai\-Wei Chang, and Dong Yu\.LongMemEval: Benchmarking chat assistants on long\-term interactive memory\.arXiv preprint arXiv:2410\.10813, 2024\.[https://arxiv\.org/abs/2410\.10813](https://arxiv.org/abs/2410.10813)\.