Retention
keeps information available across sessions.
Abstract
Persistent agent memory prevents loss of prior work, yet it can preserve conclusions after their evidence has been invalidated. Retrieving a correction does not determine which conclusions remain supported. We study selective reconsolidation, a memory operation that withdraws unsupported consequences while retaining conclusions with independent justification. The mechanism records AND/OR justification structure, re-evaluates it after review, and compiles bounded working context from the revised state.
We prove that influence-only reachability permits exact revision for every audit if and only if each claim has a unique minimal justification. In SciMapRepair, selective reconsolidation grounds 139/140 delayed answers, compared with 117/140 when the same memory records a correction without propagating it. In 80 paired worlds with identical text, timestamps, embeddings, and influence edges, justification-aware repair is exact while no propagation under-revises and influence propagation over-revises.
Introduction
Memory Beyond Retrieval
Persistence prevents forgetting while allowing contamination to endure. A research agent may retain an observation, its interpretation, a mechanistic conclusion, and the experiment planned from that conclusion. If the observation is later invalidated, storing the correction does not automatically stop the agent from retrieving and citing the derived conclusion.
Most agent-memory work asks what should be stored and retrieved. Long context retains history but remains sensitive to position and observation budget. Retrieval-augmented generation (RAG), graph retrieval, and hierarchical maps improve access under bounded context. PEEK stores an orientation cache for recurring situations.
stored state
→
relevant past information
new evidence
→
revised active belief state
We call the second operation selective reconsolidation. Retrieval relevance and revision dependence are different relations. Similarity can retrieve a correction; it cannot by itself establish which consequences lose support.
External representations reduce working-memory demand and change how a cognitive task is performed. Language-agent architectures now externalize episodic records, summaries, plans, and learned skills. This motivates separating three memory functions.
Retrieval
selects past information relevant to the current task.
Revision
changes the active consequences of stored information after its epistemic status changes.
Justification-Aware Selective Reconsolidation
A Cognitive Control Loop
The loop separates metacognitive diagnosis from memory-state revision. A discrepancy first signals that the active account no longer agrees with an executed calculation, retrieved source, or other observable result.
Selective reconsolidation changes what remains active
persistent evidence is preserved; working belief state is revised
Long-term memory
evidence, beliefs, and support handles
Discrepancy
Reviewer challenges an executed premise
Reactivation
inspect the affected supports
Reconsolidation
withdraw unsupported; keep alternatives
Working context
active beliefs with a support certificate
discrepancy→diagnosis→reactivation→source monitoring→reconsolidation→working-memory compilation
A background reviewer compares a claim with executed work and identifies a discrepancy. The controller treats the failed evidence as a reactivation root, evaluates the support families of its dependents, writes invalidated revisions only for claims with no intact support, and propagates from those newly unsupported claims. The next context excludes invalidated latest revisions and renders each active derived claim with its valid supports.
Once a root is nominated, reactivation bounds the update. The controller traverses only the downstream justification region instead of rebuilding the entire memory. Within that region, support validation requires each retained claim to have at least one intact evidential route. Reconsolidation commits new active-state revisions for claims that have lost every route while preserving their immutable pre-audit versions. Finally, working-memory compilation renders the revised latest state under a fixed context budget.
Beliefs Need Justifications
The Same Failed Source Can Produce Two Different Outcomes
Revision must also preserve unaffected beliefs. Propagating every correction through every connected node removes stale consequences, but it can also erase a claim that retains an independent justification. The memory must distinguish jointly required evidence from alternative support. A graph that records only influence edges does not record that distinction.
Same text, timestamps, embeddings, and untyped edges
the failed source is identical; only retained proof grouping differs
AND world
ab
→
claim
{a, b} is the only proof
a fails ⇒ retire claim
OR world
ab
→
claim
{a} and {b} are alternative proofs
a fails ⇒ retain claim on b
Theorem 1 · Justification-structure boundary
Influence-only reachability is correct for every audit R ⊆ E if and only if c has a unique minimal justification.
Proof
If S(c) = {S}, then U(c) = S and both rules retire the claim exactly when R ∩ S ≠ ∅. If two distinct minimal supports S₁, S₂ exist, antichain minimality gives an item e ∈ S₁ ∖ S₂. For R = {e}, graph reachability retires the claim while S₂ remains intact.
Evaluation
Four Studies
The evaluation asks four questions. Study 1 tests whether changing active memory matters beyond storing and retrieving a correction. Study 2 isolates justification semantics from graph connectivity. Study 3 measures the diagnosis–repair–behavior loop without an oracle root. Study 4 tests the mechanism on heterogeneous scientific support structures. All exact-repair endpoints are computed from program state; no model judges a reported result.
139/140grounded delayed answers
80/80paired worlds repaired exactly
20/20diagnosis episodes
20/20heterogeneous cases repaired
Changing active memory matters beyond storing and retrieving a correction
SciMapRepair
grounded delayed answers
propagated consequences repaired
Repair errors become downstream decisions
80 worlds; 240 GPT-5.6 Sol generations from identical four-field renderings
exact repaired state
grounded model decision
Diagnosis–repair–behavior loop without an oracle root
controlled 20-episode set
20/20affected premise identified
20/20conditional-repair rate
20/20delayed grounded-action rate
51.3 msmedian repair latency after diagnosis
Heterogeneous scientific support structures
exact repair without the paired templates
Selective reconsolidation grounds 139/140 answers (99.3%), compared with 117/140 (83.6%) for the identical map without dependent propagation. Both select the correct option on 139/140 questions. Support validity accounts for the difference: direct-only invalidation cites withdrawn descendants 22 times and repairs 18/40 propagated consequences; reconsolidation emits no contaminated citation and repairs 40/40. Both preserve all 20 independent branches.
Claude Sonnet 4.6 reproduces the primary 40/80, 40/80, and 80/80 pattern. GPT-5 Mini is noisier but retains the ordering: 37/80, 34/80, and 75/80.
Discussion
Under-Revision, Over-Revision, and Selective Revision
The studies expose symmetric failures. Without propagation, the agent under-revises because invalid consequences remain active. Without justification semantics, it over-revises because valid conclusions lose independent support. Selective reconsolidation withdraws precisely the claims whose recorded supports are broken.
no propagation → under-revision
influence only → over-revision
justification-aware → selective revision
A revisable belief record needs both content and a recoverable account of its justification. Retrieval keys describe when the content may be useful, while revision requires an account of why the belief remains active. That account may be a justification list, a circuit, a replayable derivation log, a typed provenance structure, or an explicit decision to abstain. The functional requirement is that the distinction remain recoverable when evidence changes.
Conclusion
Long-lived agents must revise active beliefs when corrections arrive. Retrieval can surface a discrepancy, while selective revision depends on which derivations the discrepancy breaks. Untyped influence reachability is sufficient in the unique-justification case. Once beliefs have independent support, correct reconsolidation requires justification structure or an equivalent reconstruction process.
Persistent memory should retain enough derivational information to revise beliefs selectively when their evidence changes.
Reproducibility and AI Disclosure
The anonymous supplement provides the study definitions, 1,660 model outputs, deterministic mechanism outcomes, product repair checks, statistical analyses, theory enumerations, figure sources, and a SHA-256 manifest for the reported studies. It includes the 20 SciMapRepair workspaces, 80 RevisionPair worlds, 20 diagnosis episodes, and 20 heterogeneous cases. The experiments use no human participants or private research data.
OpenAI GPT-5.6 Sol was used through Mantis Forge for literature retrieval, implementation assistance, manuscript drafting, SciMapRepair generations, RevisionPair downstream decisions, Study 3 diagnosis and downstream decisions, and Study 4 downstream decisions. OpenAI GPT-5 Mini and Anthropic Claude Sonnet 4.6 were used for the RevisionPair cross-model replication. The authors checked the proofs, inspected generated artifacts, and verified references and calculations; they take responsibility for the theory, experimental design, and paper. Primary scoring and statistical analysis are programmatic, and no model judge determines a reported endpoint.