Source-linked AI summary
CAVE-Mem: Boundary-Aware Experience Validation for Memory Search
Xinyu Li
TL;DR
Experience reuse can be harmful when a relevant lesson does not fit the current memory substrate, answer contract, or evidence boundary. CAVE-Mem gates typed operators on those conditions and cross-fitted utility, abstaining when they do not pass. Across three QA settings it improves over retrieval, deep-search, and experience-reuse baselines.
Problem
Relevant past experience may still be inappropriate when the current memory substrate, answer contract, or evidence boundary differs, so relevance alone does not establish validity.
Source: §I · p. 1Method
CAVE-Mem represents experience as typed operators and applies them only when substrate, answer-contract, evidence-boundary, and cross-fitted utility checks pass, otherwise abstaining to the base answer.
Source: §I · p. 1, F. Positive Utility · p. 3, G. Deep Search with Validated Experience · p. 4Results
Across LoCoMo, HotpotQA, and NarrativeQA, CAVE-Mem improves over retrieval, deep-search, and experience-reuse baselines; on LoCoMo with GPT-4o-mini, it gains +5.85 F1 over the strongest compared baseline.
Source: §I · p. 1, A. RQ1: Overall Performance · p. 5Takeaways & Limitations
Reusable memory should include the conditions for use, not only the instruction to reuse.
Source: §VII · p. 7
Abstract
from arXiv · showhide
Long-term memory agents increasingly rely on it- erative search and reusable experience to answer questions over large personal, factual, or narrative histories. However, current experience-memory systems largely optimize relevance: they re- trieve past search lessons that appear similar to the current state and inject them into the prompt. A relevant experience can still be harmful when the memory substrate, question intent, answer granularity, or evidence boundary changes. We propose CAVE- Mem, a training-free framework that represents experience as a typed intervention operator with applicability, boundary, and utility conditions. CAVE-Mem first obtains a base memory-search answer, then allows an operator to change it only if the oper- ator matches the current substrate, answer contract, evidence boundary, and cross-fitted utility; otherwise the system abstains. Experiments across long-term conversational memory, multi-hop question answering, and long-document narrative reasoning show consistent gains over relevance-only experience reuse.
I. INTRODUCTION
Long-term memory agents must answer questions requiring fine-grained historical information, but relevant reusable experience can still be invalid when the substrate, answer contract, or evidence boundary changes. CAVE-Mem treats reuse as a conditional intervention, applying typed operators only when validity checks pass and otherwise preserving the base answer; it improves over retrieval, deep-search, and experience-reuse baselines across three settings.
I. INTRODUCTION
- Experience reuse can fail despite semantic similarity when the memory substrate, requested answer contract, or evidence boundary differs.Examples include temporal operators misapplied to narrative prose, counting operators used for duration or frequency questions, and slot-oriented answers that drop modifiers in causal questions.Source: §I · p. 1
- CAVE-Mem is a training-free gate that lets typed experience operators affect an answer only when substrate, contract, evidence-boundary, and held-out utility checks pass.Otherwise, it abstains to the base answer; operator families include temporal-state resolution, exact-slot restoration, and answer canonicalization.Source: §I · p. 1
- Across LoCoMo, HotpotQA, and NarrativeQA, CAVE-Mem improves over retrieval, deep-search, and experience-reuse baselines.The largest gains occur on LoCoMo, where temporal and exact-slot failures are frequent.Source: §I · p. 1
- The paper formulates experience reuse as a conditional intervention decision rather than a pure retrieval problem.The approach is evaluated across long-term conversational memory, multi-hop question answering, and narrative reasoning.Source: §I · pp. 1–2
II. RELATED WORK
Prior work establishes iterative memory search and reusable agent experience, while leaving open whether retrieved guidance is valid for a current question and evidence state. CAVE-Mem focuses on this inference-time control problem rather than memory storage.
A. Iterative Deep Memory Search for LLM Agents
- Iterative deep-memory systems combine planning, retrieval, evidence integration, reflection, and reusable trajectory-level experience, raising the question of when that experience is valid for a current search state.These systems preserve raw or lightly processed history rather than compressing it into one static representation.Source: A. Iterative Deep Memory Search for… · p. 2, Figure 2 · p. 2
B. Experience Learning and Agent Self-Evolution
- Agent self-improvement work externalizes reflections, skills, or procedural knowledge, but existing experience memories mainly emphasize externalization, retrieval, and organization.CAVE-Mem studies whether retrieved guidance should intervene under the current question, evidence, and answer contract.Source: B. Experience Learning and Agent… · p. 2
C. Validity, Negative Transfer, and Process Critique Signals
- Experience may transfer harmfully beyond its validity domain, which depends on question type, memory substrate, available evidence, and expected answer granularity.CAVE-Mem addresses suppression of invalid guidance through applicability predicates, boundary vetoes, utility diagnostics, and abstention; its focus is not memory storage itself.Source: C. Validity, Negative Transfer, and… · p. 2
III. PRELIMINARY
The paper models memory search as an iterative process and represents its trajectory as a sequence of intermediate states. These states provide the substrate for distilling, retrieving, and reusing procedural experience.
A. Deep Memory Search
- Given a question and memory store, the agent repeatedly plans, searches, integrates evidence, and reflects, while maintaining temporary working memory and deciding whether to continue.The continuation decision is represented by c_t, and z_t is the temporary working memory.Source: A. Deep Memory Search · p. 2
B. Memory Search Trajectory
- A memory-search trajectory records intermediate states, each containing the current request, action, temporary memory, and retrieved evidence.Such trajectories provide material from which procedural experience can be distilled, retrieved, and reused; the central intervention question is whether retrieved experience should change the answer at a new state.Source: B. Memory Search Trajectory · p. 2
C. Experience as Intervention
CAVE-Mem treats experience as an operator that may intervene in memory search only when it is valid for the current question and search state. It estimates validity using applicability and boundary checks, cached utility, and abstention to the base answer.
C. Experience as Intervention
- Experience is operator-level procedural memory distilled from prior search trajectories, rather than merely text appended to a prompt.Operators can provide search guidance, transform working memory, or restore bounded evidence.Source: C. Experience as Intervention · p. 2
- Standard experience retrieval optimizes relevance, while CAVE-Mem treats each candidate as a typed intervention whose utility is whether it improves the base agent's task metric.Because direct counterfactual evaluation is expensive at inference, it estimates positive utility through applicability predicates, boundary checks, and cached diagnostics.Source: C. Experience as Intervention · p. 3
A. Overview of the Framework
- CAVE-Mem is an inference-time gate that checks an intervention's substrate profile, answer contract, failure boundary, and observed utility before allowing it to affect an answer.The operator body may be guidance, a working-memory transformation, or a bounded answer transformation.Source: A. Overview of the Framework · p. 3
- If no intervention satisfies the validity checks, the system abstains and returns the base answer.The validity conditions include compatible profile and contract predicates, alongside boundary and positive-utility checks.Source: A. Overview of the Framework · p. 3
B. Principled Gate Design
CAVE-Mem makes experience reuse conditional on four validity invariants and adapts operator use to the memory substrate. The same operator can be permitted for one substrate and suppressed for another.
B. Principled Gate Design
- Experience reuse requires substrate compatibility, preservation of the requested answer contract, respect for the search trajectory's evidence boundary, and positive held-out utility.The operator families instantiate these invariants: temporal-state, slot-restoration, quote/reason, and utility operators.Source: B. Principled Gate Design · p. 3
- The gate learns when an experience may intervene; it is not a universal editor that rewrites every answer.Its conditions are independent of benchmark labels or conversation identifiers.Source: B. Principled Gate Design · p. 3
C. Substrate Profile
- Date and count anchors can be reused on episodic dialogue and fact-pack memory, but are suppressed on narrative memory unless the answer contract is sufficiently narrow.This avoids unconditional transfer of episodic temporal operators to narrative reasoning.Source: C. Substrate Profile · p. 3
D. Answer Contract
CAVE-Mem checks whether experience preserves the question's answer contract and evidence boundary, and whether its operator family has positive cross-fitted utility. Its deep-search policy selects only valid candidates and otherwise falls back to the base answer.
D. Answer Contract
- Direct slot questions often call for concise answers, whereas explanation questions require reason or event phrases and are more vulnerable to over-compression.A narrower operator is accepted only when anchored in the base answer, supported across independent candidate paths, and consistent with question type.Source: D. Answer Contract · p. 3
E. Failure Boundary
- The gate rejects candidates that alter the requested semantic object, such as dropping essential modifiers, losing temporal relations, or replacing a reason or quote with an unsupported paraphrase.Uncertain candidates are not applied.Source: E. Failure Boundary · p. 3
F. Positive Utility
- Cached operator utility is estimated from diagnostic trajectories outside the target conversation or evaluation block, and must be positive for the current substrate and answer contract.The gate does not query the target example's gold answer, correction outcome, or conversation-level aggregate score when deciding whether to change it.Source: F. Positive Utility · p. 3
G. Deep Search with Validated Experience
- Online, the base search produces a trajectory, retrieved evidence, and answer before the gate selects among typed interventions or abstains.Candidate sources include the base-answer fallback, contract operators, evidence-boundary operators, and cross-fitted utility diagnostics; the figure depicts the Profile, Contract, Boundary, and Utility checks.Source: G. Deep Search with Validated Experience · pp. 3–4, Figure 3 · p. 4
H. Abstention as a First-Class Action
CAVE-Mem treats abstention as a normal outcome when an experience candidate fails its conditions. The evaluation asks whether validated reuse beats baselines, scales across Qwen backbones, and benefits from its validation components.
H. Abstention as a First-Class Action
- When a relevant candidate's conditions are unmet, CAVE-Mem abstains to the base answer rather than letting retrieved experience change the prompt.Abstention can occur for an incompatible substrate, a broad answer contract, or a family with negative measured utility.Source: H. Abstention as a First-Class Action · p. 4
A. Research Questions
- The evaluation asks whether validated experience reuse outperforms retrieval, deep-search, and experience-reuse baselines.It also asks whether effectiveness persists across smaller and stronger Qwen backbones and how validation components and abstention relate to gains.Source: A. Research Questions · p. 4
- The ablation research question specifically examines how abstention prevents negative transfer.The other evaluation questions concern overall performance and scaling across Qwen backbones.Source: A. Research Questions · p. 4
B. Datasets and Metrics
The evaluation covers long-term conversational memory, multi-hop QA, and long-document narrative reading, with metrics chosen by dataset. LoCoMo uses a cross-fitted policy, while HotpotQA and NarrativeQA reuse frozen LoCoMo schemas and thresholds.
B. Datasets and Metrics
- LoCoMo uses nine held-out conversations, while HotpotQA covers 56K, 224K, and 448K context regimes and NarrativeQA uses a 300-question subset averaging about 87K tokens.LoCoMo is scored with token F1 and BLEU-1; HotpotQA and NarrativeQA use maximum token F1 over gold answers.Source: B. Datasets and Metrics · pp. 4–5
C. Leakage-Controlled Protocol
- LoCoMo utility thresholds are cross-fitted across held-out conversations, and target-conversation gold answers, corrections, and aggregate scores do not determine interventions.HotpotQA and NarrativeQA reuse LoCoMo operator schemas, vetoes, and threshold range without dataset-specific utility tuning or error-analysis edits.Source: C. Leakage-Controlled Protocol · p. 5
D. Baselines
- Baselines include RAG, GAM, and R2Mem, plus matched-backbone reproductions of structured-memory systems including Mem0, A-Mem, MemoryOS, LightMem, and Memory R1.RAG uses fixed 2048-token chunks, top-5 embedding retrieval, and one-pass answer generation.Source: D. Baselines · p. 5
E. Implementation Details
CAVE-Mem is evaluated with matched datasets, metrics, retrieval interfaces, and scoring scripts. It leads LoCoMo under both reported backbones and achieves the best scores on the supplied cross-dataset checks with frozen settings.
E. Implementation Details
- The main cross-dataset runs use GPT-4o-mini and text-embedding-3-small, with shared loaders, metrics, retrieval interfaces, and scoring scripts.Qwen2.5 LoCoMo runs use local vLLM endpoints and BGE-M3 embeddings.Source: E. Implementation Details · p. 5
A. RQ1: Overall Performance
- +5.85 F1 and +5.65 BLEU-1 over the strongest compared baseline with GPT-4o-mini; Qwen2.5-7B gains are +1.87 F1 and +1.59 BLEU-1.The LoCoMo results cover the 1,388-question evaluation split.Source: A. RQ1: Overall Performance · p. 5, Table II · p. 6
- CAVE-Mem achieves the best score on NarrativeQA and all three HotpotQA context regimes under the frozen schema and threshold setting.At 448K, it preserves the strongest base branch when intervention conditions are not met and applies bounded answer-contract edits only when checks pass.Source: A. RQ1: Overall Performance · pp. 5–6, Table III · p. 6
B. RQ2: Scaling Behavior
CAVE-Mem retains its LoCoMo gains under Qwen2.5 scaling while changing only a limited number of answers in several settings. The ablations indicate that candidate generation, boundary checks, and utility selection serve distinct roles.
B. RQ2: Scaling Behavior
- 34.65 overall F1 for CAVE-Mem under Qwen2.5-3B is 1.76 F1 above the strongest baseline; under Qwen2.5-7B, the gain is 1.87 F1.The same gate and operator schemas are used across backbones.Source: B. RQ2: Scaling Behavior · p. 6
C. Intervention Frequency
- On HotpotQA, the gate changes zero answers at 56K, 11 at 224K, and five at 448K; on NarrativeQA, it changes 16 of 300 answers as F1 rises from 38.73 to 39.37.On LoCoMo, answer-contract changes stay below 10% in each category, temporal validation changes 2.1% of temporal questions, and open-inference validation changes 14.5% of open-domain questions.Source: C. Intervention Frequency · p. 6
D. Failure Patterns
- LoCoMo gains are strongest on answer-contract mismatches, including off-slot attributes, lost temporal relations, dropped reason or quote components, and unsupported broad count or date overrides.Broad candidate auditing and raw-localizer overlays are diagnostic negative controls, not eligible interventions.Source: D. Failure Patterns · p. 6
E. RQ3: Ablation Study
- Ablations assign candidate generation to arbitration and exact-slot rescue, boundary checks to filtering inapplicable edits, and utility selection to deciding whether an intervention is worthwhile.Threshold sweeps show a stable plateau near the selected operating range; leave-one-family diagnostics find answer-contract checks matter most for multi- and single-hop gains, temporal checks for temporal questions, and open-inference checks for open-domain questions.Source: E. RQ3: Ablation Study · p. 6, Figure 4 · p. 7, Figure 5 · p. 7
F. Why Flat Experience Can Regress
Flat experience reuse can regress when a relevant operator conflicts with the current answer contract. CAVE-Mem conditions intervention, while corrected errors cluster around answer-slot and granularity mismatches.
F. Why Flat Experience Can Regress
- A semantically relevant operator can still hurt when its guidance conflicts with the question’s answer contract, context regime, or evidence state.Examples include latest-event guidance hurting date-constrained questions and counting guidance hurting duration or frequency questions.Source: F. Why Flat Experience Can Regress · p. 7
G. Qualitative Failure Taxonomy
- Corrected LoCoMo failures concentrate on answer-slot and granularity errors, followed by temporal, quote/feeling, and open-inference errors.The six classes include off-slot substitution, temporal drift, count looseness, quote and feeling loss, open-domain entity confusion, and over-compression.Source: G. Qualitative Failure Taxonomy · p. 7
VII. CONCLUSION
- CAVE-Mem improves over retrieval, deep-search, and experience-reuse baselines across LoCoMo, HotpotQA, and NarrativeQA.The conclusion recommends storing conditions for reuse, not only the instruction to reuse.Source: §VII · p. 7