PaperlayerView exact revision on arXiv

Source-linked AI summary

Memory Is a Derivation: The Distributed-Evidence Paradox in Long-Term Agents

Hongjun Liu, Chen Zhao

arXiv:2609.36130v1cs.AI

TL;DR

Long-running agents reuse persistent memories, but it can be unclear whether a memory follows from the interaction history available when written. DERIVAUDIT audits evidence scope, full-memory meaning, and write-time admission; broader history recovers support for many citation-insufficient memories, yet admission remains unreliable.

  • Problem

    Persistent memories may omit relevant historical evidence or combine supported facts into meaning the interaction history does not establish.

    Source: §1 · pp. 1–2

  • Method

    DERIVAUDIT compares writer citations with broader pre-write history, checks support for each memory's full meaning, and evaluates write-time admission.

    Source: §1 · p. 2

  • Results

    Nearly 60% of citation-insufficient memories regain support with broader history, but roughly one fifth remain unsupported after expansion and admission remains inconsistent across verification models.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Distorted memories can propagate incorrect information, while rejecting valid memories can remove information needed for later tasks.

    Source: §1 · p. 2

    Evidence expansion is retrieval-bounded, so failure to recover support does not establish historical invalidity.

    Source: §5 · p. 9

Abstract

from arXiv · show

Long-running LLM agents compress past interactions into persistent memories that may be reused as premises for later tasks. This creates a distinct derivation problem: whether the memory actually follows from what the interaction history supports. Relevant evidence may be scattered across earlier interactions, while compression can introduce relations or event status that the history never established. A valid memory may therefore appear unsupported because its citations omit relevant evidence, while individually supported facts may be composed into a stronger statement the history never established. We characterize this problem through three coupled requirements: (1) Evidence scope; (2) Compositional validity; (3) Admission reliability. We therefore ask whether the interaction history available at write time supports what enters persistent memory. We introduce DerivAudit, a framework for auditing whether a memory is actually supported by the history available when it was written. The audit separates three questions: whether supporting evidence lies beyond writer-provided citations, whether the composed memory introduces unsupported meaning, and how write-time admission decisions affect later memory use. Across two natural memory corpora, audits using broader pre-write history recover support for nearly 60% of memories that appear unsupported from citations alone, while 17-21% remain unsupported after expansion. Yet broader evidence does not by itself make admission reliable: unsupported memories are still frequently admitted across verification models, and evidence expansion alone worsens it on two backbones.

1 INTRODUCTION

Persistent memory can become a premise for later tasks, yet a write may misrepresent what interaction history supports. DERIVAUDIT frames this as a derivation problem and audits evidence scope, compositional validity, and admission reliability.

1 INTRODUCTION

  • A memory may be valid despite incomplete citations, or combine supported facts into a relation or event status the history never established.Later tasks may reuse stored information as established fact, even when the original interaction is no longer in context.Source: §1 · pp. 1–2
  • DERIVAUDIT audits whether interaction history available at write time supports persistent memory, including its full meaning and write-time admission.It compares writer-provided citations with broader pre-write history and checks support beyond individual facts.Source: §1 · p. 2
  • Nearly 60% of citation-insufficient memories regain support from broader history, while roughly one fifth remain unsupported after expansion.Across the evaluated corpora, admission remains inconsistent across verification models.Source: §1 · p. 2
  • The paper frames memory writing as derivation, provides a common audit setup, and studies incomplete provenance, unsupported composition, and verification-induced rejection of valid memories.It also examines how distorted and missing memories can affect downstream behavior.Source: §1 · pp. 2–3

2 RELATED WORK

The related-work discussion places DERIVAUDIT across long-term agent memory, reliable memory writing, and factuality verification. Its focus is whether interaction history justified a memory when it entered persistent state.

2 RELATED WORK

  • DERIVAUDIT shifts the target from whether stored information can later be recalled to whether memory was justified when it entered persistent state.The paper situates this distinction within research on long-term agent memory and its operations.Source: §2 · p. 3
  • The setting highlights two challenges: writer-provided provenance may omit supporting history, and individually supported facts may not license composed relations.The comparison concerns evidence beyond attached provenance, compositional meaning, and consequences of admission decisions.Source: §2 · p. 3, Table 1 · p. 3
  • DERIVAUDIT extends factuality and compositional verification to longitudinal interaction histories whose derived memories may later be reused without their original sources.It jointly examines where support resides, what composition adds, and which memories survive verification and affect later behavior.Source: §2 · p. 3

3 DERIVAUDIT: AUDITING MEMORY AS DERIVATION

DERIVAUDIT audits whether persistent memories follow from the interaction history available when they are written. It separates evidence scope, compositional validity, and admission decisions that shape later memory use.

3 DERIVAUDIT: AUDITING MEMORY AS DERIVATION

  • The audit treats agent memory as a compact record distilled from earlier interactions and asks whether history supports its full meaning and later admission.It separates where support lies, whether the full memory is supported, and whether verification retains the right information.Source: §3 · p. 4

3.1 EVIDENCE SCOPE: RECOVERING THE HISTORY BEHIND A MEMORY

  • Up to 12 BM25-retrieved pre-write passages supplement writer citations, within a total evidence budget of 16 passages.Later interactions and previously written memories are excluded; insufficient support after retrieval-bounded expansion remains unresolved rather than proving historical absence.Source: §3.1 · p. 4

3.2 COMPOSITIONAL VALIDITY: CHECKING WHAT THE AGENT WILL REMEMBER

  • A memory is considered supported only when evidence supports every semantic obligation, including its claims, relations, and qualifications.For example, a causal statement requires support for both events and their causal relation.Source: §3.2 · p. 4
  • Verification compares holistic factuality, atomic-claim checks, and predicate–argument QA with views that represent relations and qualifications.The comparison tests whether each representation preserves the compositional meaning needed to judge support.Source: §3.2 · pp. 4–5
  • Across 70 matched example families, controlled conditions distinguish integrating distributed evidence from rejecting unsupported compositions.Valid distributed cases require combining evidence across passages; invalid cases pair plausible premises without support for their relation.Source: §3.2 · p. 5

3.3 ADMISSION RELIABILITY: RETAINING KNOWLEDGE FOR LATER TASKS

  • Three write-time settings replay the same candidate writes, while the two expanded-history settings hold evidence constant to distinguish evidence scope from verification representation.The settings are citation-only holistic, expanded-history holistic, and expanded-history obligation-aware verification.Source: §3.3 · p. 5
  • Adding two or four independently true obligations tests whether more required checks reduce verification scores or admission of supported memories.The candidate, evidence, and reference validity remain unchanged; the authors call the effect accumulated verification noise.Source: §3.3 · p. 5
  • Downstream probes compare correct, distorted, and absent memories under a fixed task context to isolate carrying an incorrect premise from removing needed information.The study also tracks whether later rewriting preserves timing and uncertainty qualifications and whether stored memories are retrieved.Source: §3.3 · p. 5

4 EXPERIMENTS

The experiments audit natural and controlled memory writes across evidence scope, compositional validity, and admission reliability. Broader history often recovers support missed by citations, but unsupported compositions remain difficult to reject and detailed checking can also reduce valid-memory retention.

4.1 WHAT DOES WRITE-TIME VERIFICATION RETAIN AND REJECT?

  • On Qwen, expanded history raises provenance-repaired retention from 82.1% to 93.8%, while unsupported-memory admission remains 80.8%.Across models, obligation-aware checking lowers unsupported admission on Gemma from 66.7% to 59.0% but raises it on Llama from 83.3% to 88.5%.Source: §4.1 · p. 6, Table 2 · p. 7

4.2 WHY IS MEMORY ADMISSION DIFFICULT?

  • Citations alone are insufficient for 42.5% of LoCoMo and 51.0% of HaluMem-Medium memories; expanded history recovers support for about 60% of those cases.The audit includes 400 unedited writes and controlled evaluations that vary evidence distribution while holding candidate memory and validity fixed.Source: §4 · p. 6, §4.2 · p. 7, Figure 3 · p. 8
  • On Qwen, compact-relation verification detects 59.5% of unsupported relations while retaining 94.9% of valid multi-passage memories; distributed evidence can make invalid compositions harder to reject.With distributed evidence, holistic and obligation-aware verification detect only 8.6% and 11.4% of invalid Qwen cases, respectively.Source: §4.2 · p. 7, Table 3 · p. 7, Figure 3 · p. 8

4.3 WHAT HAPPENS AFTER THE WRITE-TIME DECISION?

  • Supplying distorted memories raises downstream failure by 93.0, 83.8, and 92.2 percentage points for Gemma, Qwen, and Llama; removing valid memories also increases failure.Natural retrieval produces misinformation in 100% and 96.2% of invalid-memory cases for Gemma and Qwen, respectively.Source: Figure 4 · p. 8, §4.3 · p. 9

4.2 WHY IS MEMORY ADMISSION DIFFICULT?

  • Four additional true obligations lower Qwen’s median support score by 7.7 log-odds units, while admission falls to 2.6% on Gemma and 35.5% on Llama.The authors call this accumulated verification noise; increasingly detailed checking can make valid memories harder to retain.Source: §4.2 · p. 8

4.3 WHAT HAPPENS AFTER THE WRITE-TIME DECISION?

  • By round 4, temporal and modality drift reach 69.7% and 78.1% for Gemma, and 91.1% and 92.6% for Qwen.These trends are descriptive, since citation breadth, writer settings, and record populations vary across rounds.Source: §4.3 · p. 8
  • In a natural write, a future plan becomes an ongoing activity, and general social support becomes support for that specific effort.Neither stronger claim about Sharon Brown’s sports academy is established by the pre-write history.Source: Figure 5 · p. 9, §4.3 · p. 9

5 CONCLUSION AND LIMITATIONS

The paper concludes that reliable memory writing requires checking evidence scope, whether evidence supports a memory’s full meaning, and which memories verification admits. Its findings are bounded by retrieval, labeling, evaluation, and system limitations.

5 CONCLUSION AND LIMITATIONS

  • Nearly 60% of citation-insufficient memories regain support with broader history, but unsupported writes remain and admission reliability varies across verification models.The authors also report that both distorted and missing memories can substantially affect later answers.Source: §5 · p. 9
  • Reliable memory writing requires more than finding relevant evidence: verification must also determine whether the evidence supports the full meaning entering memory.The paper identifies evidence scope, compositional validity, and admission reliability as coupled challenges.Source: §5 · p. 9
  • Findings are limited to evaluated settings: expansion is retrieval-bounded, natural and controlled data have coverage limits, and results vary across verification choices.Reference labels rely primarily on two model assessments with blinded adjudication; stratified human review showed 96.0% agreement. Consolidation results are observational, and downstream interventions do not capture a fully closed-loop memory system.Source: §5 · p. 9

A FORMAL BOUNDARIES AND PROOFS

The appendix states boundaries on what evidence and representations can support verification, then describes primitive-based annotation and adjudication procedures.

A.1 EVIDENCE AND INFORMATION BOUNDARIES

  • A maximum over individual evidence scores rejects a memory when each span falls below threshold, even if the spans jointly license it.The boundary applies to independent-span scoring, not verifiers that score evidence subsets jointly.Source: §A.1 · p. 13
  • If an interface deletes the only differing relation and gives two candidates identical inputs, no downstream scorer can recover that relation.This boundary concerns information loss, not verifier architecture.Source: §A.1 · p. 13
  • The boundaries do not rule out scalar or holistic verification; failures arise from insufficient evidence scope or omitted truth-conditional components, not a single score itself.A scalar score can be semantically complete by construction.Source: §A.1 · p. 13
  • Repair distinguishes logical weakening, which must be entailed by the original assertion, from evidence-grounded correction, which uses a separately history-licensed replacement.COMPLETED(p) does not generally entail PLANNED(p), so that edit needs another premise to count as logical weakening.Source: §A.1 · p. 13

B ANNOTATION INSTRUMENT AND ADJUDICATION

  • Reference labels are routed from three ordered primitives: determinate truth conditions, evidence entailing the obligation, and evidence entailing its negation.Support and refutation use the same obligation and all spans, differing only in direction; the route maps primitive combinations to four labels.Source: B ANNOTATION INSTRUMENT AND ADJUDICATION · pp. 13–14
  • Routing showed 0.81–0.91 agreement stability under rubric revisions, compared with 0.69–0.78 for directly typed multi-way labels.Under routing, rubric wording moves the primitives, and labels move only as far as the primitives do.Source: B ANNOTATION INSTRUMENT AND ADJUDICATION · p. 14
  • Refutation requires the same subject, event, and time frame; a later state change counts as an update, not a contradiction.The frozen rule applies identically in both evidence scopes.Source: B ANNOTATION INSTRUMENT AND ADJUDICATION · p. 14
  • Two independent model instances annotate each record, disagreements alone receive blinded adjudication, and record verdicts follow contradicted > unsupported > ambiguous > supported precedence.Clerical defects were repaired through fresh-instance re-annotation rather than editing labels.Source: B ANNOTATION INSTRUMENT AND ADJUDICATION · p. 14

C NATURAL-AUDIT ACCOUNTING AND PILOT GATE

The section accounts for natural-audit labels and describes how evidence scopes, retrieval limits, and prevalence estimates are handled.

C NATURAL-AUDIT ACCOUNTING AND PILOT GATE

  • Across records annotated by both, 643 judgements would exist if none were spurious; removing 20 marks—16 from 8 dropped obligations and 4 from one-annotator marks—leaves 631 support-status-family units.The drops close exactly: 8×2+4=20 spurious marks, and 643−8−4=631 units.Source: C NATURAL-AUDIT ACCOUNTING AND PILOT… · p. 14
  • Post-adjudication, 4 of 631 final labels were contradicted (0.63%); prevalence uses that count, while reliability coefficients are pre-adjudication.The section reports four contradicted-label denominators because this label is rare.Source: C NATURAL-AUDIT ACCOUNTING AND PILOT… · p. 14
  • The pilot gate demanded all four labels in 12 records despite an expected contradicted count of 0.72 and about a 0.487 chance of none; rare-label results remain descriptive.The authors record the gate as a protocol deviation and note that occupancy checks must be powered for the rarest demanded label.Source: C NATURAL-AUDIT ACCOUNTING AND PILOT… · pp. 14–15

D EVIDENCE SCOPES AND RETRIEVAL CONFIGURATION

  • Each natural record is audited with writer-cited spans and again with those spans plus up to 12 retrieved pre-write passages, capped at 16 total.The two views are unlinkably identified and spans are presented chronologically without marking cited versus retrieved evidence.Source: D EVIDENCE SCOPES AND RETRIEVAL… · p. 15
  • Search-cap records with expanded-scope insufficient verdicts are marked unresolved, while prevalence is conversation-balanced and paired scope contrasts account for record clustering.The write-weighted alternative moves reported rates by at most 0.017 and is reported alongside.Source: D EVIDENCE SCOPES AND RETRIEVAL… · p. 15

E CONTROLLED SUITE CONSTRUCTION

The controlled suite matches memory families and manipulates evidence placement and licensing, while equal-budget prompt search tests verification interfaces.

E CONTROLLED SUITE CONSTRUCTION

  • Seventy validated seed families instantiate local or distributed evidence with valid or invalid candidate memories, matched on semantic type, obligation count, length, and distractors.Truth is set by construction, and prompt files are leak-checked mechanically.Source: E CONTROLLED SUITE CONSTRUCTION · p. 15
  • Interventions remove all sufficient support, remove only redundant evidence, or add a temporally applicable refutation; a control replaces plausible premises with matched irrelevant content.One family was dropped from the mechanism control for residual content overlap, leaving 69.Source: E CONTROLLED SUITE CONSTRUCTION · p. 15
  • The faithful-atomic interface’s best of nine configurations remains below other interfaces’ two-shot means, so its decomposition deficit persists across its prompt-search budget.Each interface receives three paraphrase templates crossed with 0, 2, or 4 few-shot examples; the winning configuration is selected on development data and evaluated on untouched test conversations.Source: E CONTROLLED SUITE CONSTRUCTION · p. 15

F THRESHOLD SELECTION AND DECISION RULES

The section audits verdict scoring and sets backbone-specific decision rules. It also reports how uncertainty and rerun variation are handled.

F THRESHOLD SELECTION AND DECISION RULES

  • Verdicts come from token log-probabilities; analyses use log-odds, and records missing both verdict tokens are counted as scoring errors.The serving window is widened to the top 100 log-probabilities, and requests are rejected rather than silently truncated if that window cannot be honoured.Source: F THRESHOLD SELECTION AND DECISION RULES · p. 16
  • 0 of 120 Gemma decisions and 1 of 33 Qwen decisions flipped after rail-floor scores were replaced with exact probabilities.The Qwen flip was a +0.031-margin record within known rerun variation.Source: F THRESHOLD SELECTION AND DECISION RULES · p. 16
  • Qwen uses cross-fitted log-odds thresholds, while Gemma and Llama use their untuned argmax operating points.At argmax, Qwen falsely admits 96–99% of natural records, whereas the dense backbones already gate; comparisons stay within each backbone.Source: F THRESHOLD SELECTION AND DECISION RULES · p. 16
  • 12:6 discordant pairs on 78 negative records leave the admission reduction directionally consistent but unresolved.Policy contrasts use exact McNemar tests and record-paired and conversation-clustered bootstrap intervals; the clustered intervals are quoted in the main text.Source: F THRESHOLD SELECTION AND DECISION RULES · p. 16
  • Median absolute rerun log-odds differences are ≈0.2–0.5 on the mixture-of-experts backbone and ≈10−8 on dense backbones.The Qwen rail-audit flip falls within the recorded variation.Source: F THRESHOLD SELECTION AND DECISION RULES · p. 16

G ROBUSTNESS AND PROTOCOL CONTROLS

Robustness and protocol controls change verification outcomes or operating points, but none supplies a universal repair. The controls also isolate plausible-premise composition bias as the mechanism behind the distributed-evidence paradox.

G ROBUSTNESS AND PROTOCOL CONTROLS

  • Representation structure helps or hurts depending on backbone and rendering, and no representation is a universal repair.Table 4 compares unsupported-relation detection with multi-span valid retention across correct, isomorphic, shuffled, filler, and compact-relation renderings.Source: G ROBUSTNESS AND PROTOCOL CONTROLS · p. 16, Table 4 · p. 17
  • Equal-budget prompt search brings the top three interfaces within four points, but the winning interface flips across backbones.Alternative aggregation changes operating points without removing the tension; on Qwen, minimum, mean, and noisy-AND yield 0.266, 0.101, and 0.304 unsupported detection, respectively.Source: G ROBUSTNESS AND PROTOCOL CONTROLS · p. 16, Table 5 · p. 17
  • Adding independently verified true obligations lowers support scores monotonically on every backbone, while the mechanism controls isolate plausible-premise composition bias.The paradox disappears when plausible premises are replaced with token-matched irrelevant content; the verifier notices in-input support deletion and honours applicable corrections.Source: G ROBUSTNESS AND PROTOCOL CONTROLS · p. 16
  • Llama derivation-aware representation variants were not run in this suite.Gemma and Llama use their argmax operating points.Source: G ROBUSTNESS AND PROTOCOL CONTROLS · p. 17
  • Oracle-input results use a separate 235-member minimal-pair suite and should not be numerically compared with the 512-record main verification matrix.Equal-budget prompt search uses the same nine-configuration budget for each interface.Source: G ROBUSTNESS AND PROTOCOL CONTROLS · p. 17

H DOWNSTREAM CAUSAL PROTOCOL

The downstream protocol measures the effects of distorted or absent memories while controlling the reader’s retrieval context and response quality. Its results show two-sided consequences across backbones, alongside controls for accumulated verification noise and the composition mechanism.

H DOWNSTREAM CAUSAL PROTOCOL

  • The probe changes one memory slot to correct, distorted, or absent while holding retrieval position and surrounding context fixed.It measures CACEharm on a harm query and CACEreject on a content query, paired within example; all six answer-option permutations are averaged.Source: H DOWNSTREAM CAUSAL PROTOCOL · p. 17
  • All backbones pass four all-or-nothing control floors at 0.80 before any causal quantity is read.The floors test correct answering with the record present, recognition of absence, correct recall, and not manufacturing reasons.Source: H DOWNSTREAM CAUSAL PROTOCOL · p. 17
  • Adding independently verified true obligations lowers support scores on every backbone, with backbone-scaled magnitude, although binary acceptance moves little on the graded backbone.The candidate memory stays unchanged; the k=2 filler list is a prefix of k=4, isolating the number of added obligations.Source: Table 6 · p. 18, H DOWNSTREAM CAUSAL PROTOCOL · p. 18
  • Table 7’s controls show that the verifier notices in-input support deletion and applicable corrections, while plausible-premise composition bias disappears with token-matched irrelevant content.Median paired ∆log-odds are 28–37 for full support deletion, 0.1–1.4 for redundant-path deletion, and 32–37 for explicit correction; Gemma’s applicable-correction false-admit rate is 0.043–0.145.Source: Table 7 · p. 18, H DOWNSTREAM CAUSAL PROTOCOL · p. 18
  • CACEharm is +0.838 (Qwen), +0.930 (Gemma), and +0.922 (Llama); CACEreject is +0.928, +0.995, and +0.949, respectively.The section reports persistent two-sided consequences across backbones, not a property confined to one model class.Source: H DOWNSTREAM CAUSAL PROTOCOL · p. 18

I ADDITIONAL NATURAL CASES

J VERIFICATION INTERFACE PROMPTS

  • The whole-record prompt requires support for every claim and relation, while obligation-aware joint verification checks the listed obligations together.The interfaces share evidence rendering and verdict extraction; multi-unit interfaces aggregate unit scores by the minimum.Source: J VERIFICATION INTERFACE PROMPTS · pp. 18–20
  • Predicate–argument QA scores derived questions against the full evidence block, either jointly or per question with minimum aggregation.The atomic-claim and residual prompts separately test listed claims and what the record asserts beyond them.Source: J VERIFICATION INTERFACE PROMPTS · pp. 19–20

K IMPLEMENTATION AND REPORTING

The implementation section records experiment configurations and sets a reporting checklist for result tables, including separate reporting of ranking measures and score-saturation audits.

K IMPLEMENTATION AND REPORTING

  • Table 9 records writer, verifier, and reader configurations, serving parameters, and evidence budgets for each experiment.It summarizes evaluation suites for evidence scope, compositional validity, admission reliability, and later memory use.Source: K IMPLEMENTATION AND REPORTING · p. 20, Table 9 · p. 21
  • Result tables must report sampling, splits, thresholds, models, retrieval, uncertainty, and operational details where applicable.Within-pair ranking and standard ROC-AUC are reported separately, alongside unclipped label log-probability differences, unique-score counts, and pairwise ties.Source: K IMPLEMENTATION AND REPORTING · p. 20
Loading 2609.36130v1…