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.
Figures & tables
Fig. 1: From memory compression to flat experience reuse and validated intervention. CAVE-Mem validates a candidate experience before it can change the base answer; otherwise, it abstains and preserves the base answer.
Fig. 2: Deep memory search iterates over planning, memory search, integration, and reflection, producing retrieved evidence rt and a state trajectory s0,…,sT .
Fig. 3: CAVE-Mem framework. Offline diagnostics produce cached trajectories, candidate families, a utility ledger, and a veto map. Online, the base memory-search agent produces retrieved evidence and a base answer, while candidate interventions are routed through the CAVE-Mem validity-gate controller/shield. The gate checks Profile, Contract, Boundary, and Utility before choosing between applying the candidate and abstaining to the base answer.
Gate check
Valid intervention
Boundary veto
Role in CAVE-Mem
Substrate profile
Operator family is enabled for the current memory substrate, e.g., episodic dialogue, fact-pack QA, or narrative text.
Operator transfers a substrate-specific procedure outside its profiled domain.
Prevents cross-substrate transfer.
Answer contract
Candidate directly fills the requested slot, e.g., person, place, date, count, or object.
Candidate answers a nearby attribute or compresses a reason into an unsupported phrase.
Aligns answer shape with the question.
Temporal boundary
Candidate preserves the exact date or state transition anchored in retrieved evidence.
Candidate replaces a constrained date with a broader month/year or an unsupported current state.
Prevents temporal drift.
Count boundary
Candidate is backed by enumerated instances with stable wording.
Candidate uses vague forms such as “at least” when an exact count is required.
Converts evidence into exact counts.
Quote/reason boundary
Candidate preserves source-grounded reason, feeling, or quoted wording.
Candidate paraphrases away the answer-bearing phrase.
Protects short answer-bearing spans.
Utility ledger
Operator family has positive cross-fitted diagnostic effect under the current substrate.
Candidate source is high-variance under matched counterfactual replay.
Selects high-utility interventions.
TABLE I: Gate checks used by CAVE-Mem to convert retrieved experience into validated interventions.
Model
Method
Overall
Multi-hop
Temporal
Open
Single-hop
F1
BLEU
F1
BLEU
F1
BLEU
F1
BLEU
F1
BLEU
GPT-4o-mini
RAG [ 5 ]
39.92
34.52
33.75
24.33
31.52
26.86
25.21
20.42
46.60
42.16
Mem0 [ 11 ]
48.27
42.10
30.73
20.91
52.14
46.13
31.56
25.70
54.33
49.25
A-Mem [ 12 ]
46.59
40.55
32.34
23.77
49.34
43.20
30.33
25.00
51.94
46.70
MemoryOS [ 13 ]
45.26
39.28
29.17
20.00
52.67
46.62
30.46
24.71
49.35
44.40
LightMem
45.80
39.90
29.01
20.48
47.50
41.92
26.71
21.53
52.67
47.43
TABLE II: LoCoMo results under GPT-4o-mini and Qwen2.5-7B-Instruct. We report token F1 and BLEU-1 on the 1,388-question evaluation split.
Method
NQA 87K
HP 56K
HP 224K
HP 448K
RAG [ 5 ]
32.54
46.86
31.52
25.98
GAM [ 14 ]
38.66
64.07
60.19
59.36
R2Mem [ 15 ]
37.24
62.74
61.27
61.70
CAVE-Mem (ours)
39.37
66.24
61.58
61.87
TABLE III: Cross-dataset long-context QA check under GPT-4o-mini. NarrativeQA uses the 300-question average-87K subset; HotpotQA reports the 56K, 224K, and 448K context regimes.
Fig. 4: Category-level component ablation on LoCoMo. Bars correspond to five diagnostic checkpoints: profile, ELAA arbitration, validity-boundary checking, exact-slot rescue, and utility-aware selection. Conceptually, ELAA arbitration and exact-slot rescue are two candidate-intervention generators within the same high-level component.
Fig. 5: Sensitivity analysis for utility-gated intervention on LoCoMo. Each point uses conversation-level cross-fitting; the shaded region marks the pre-specified operating range rather than a per-conversation optimum.
Fig. 6: Leave-one-boundary-family ablation on LoCoMo. Each panel reports the score drop from full CAVE-Mem after disabling one validation family.
Large language model (LLM) agents now autonomously complete long-horizon tasks, such as repository-level coding, through hundreds of think-act steps, yet the experience each run produces is rarely reused on the next. Prior memory systems concentrate on retrieval, finding the most relevant past trajectory. Retrieval cannot observe the running agent, so it cannot decide whether the agent needs an experience, at which point, or for how long. We introduce REM (Reusable Experience Memory), a framework that manages the memory lifecycle from inside the agent's think-act loop through three explicit decisions. What to store: each completed trajectory is consolidated into a phase-structured document, two orders of magnitude smaller than the trajectory. When to recall: the document is injected at the start of the task; the costlier implementation record is withheld until an in-loop gate detects the agent struggling. What to forget: the record is evicted after a single exposure. On repository-level bug fixing (SWE-ContextBench Lite), REM retrieves the related trajectory more accurately than every state-of-the-art memory system (68.7% vs. 55.6% match@1) and improves, in every replicate, end-task resolution (27.6% vs. 24.9%) at 17% lower cost. Injecting the same experience as a summary or raw trajectory, even when oracle-retrieved, does not. Ablating any of the three decisions reduces resolution. Altogether, these results demonstrate that the bottleneck in experience reuse is not retrieval fidelity but how experience is stored and governed across its lifecycle, opening a control-theoretic perspective on agent memory.
Xinyi Lisa Chen, Pedro Herrero-Vidal, Xiaohan Chen +6
Long-term language-model agents accumulate memories across interactions, but their retrievers typically do not accumulate retrieval experience. Semantic retrieval is efficient, but embedding similarity does not always reflect whether a memory contains evidence relevant to the current query. Large language model (LLM) rerankers provide stronger query-conditioned relevance scores, yet stateless reranking repeatedly scores a large candidate pool and discards these scores after each query. We introduce EARM, an experience-amortized reranking framework that treats previously acquired LLM relevance scores as reusable retrieval experience. EARM stores sparse query--memory relevance scores in an online matrix, learns their shared structure through causal matrix completion, and combines a small set of newly observed scores with estimated scores to rerank the remaining candidates. The scoring budget decreases as experience accumulates, changing LLM reranking from a repeated per-query expense into a retrieval capability learned over an agent's lifetime. Experiments on long-term conversational memory show that mixed observed-and-estimated reranking improves answer accuracy over semantic retrieval by up to 6.62% and remains effective when only 17.5% of candidates receive direct LLM relevance scores, thereby substantially reducing the inference overhead of LLM reranking. These results motivate a broader view of agent memory: a long-lived agent should remember not only past content, but also how that content has proved useful for retrieval.
Qi Feng, Chris Ding, Jicong Fan
School of Data Science The Chinese University of Hong Kong, Shenzhen
Long-term memory has become increasingly important for LLM agents that operate across extended interactions and evolving task contexts. Recent memory systems have made past experiences more persistent, compact, and retrievable, but retrieval alone does not ensure that a memory provides valid evidence for the current query. When experiences are compressed into reusable fragments, memories from different situations may appear equally relevant if they involve recurring entities or user states. We refer to this failure as context collapse: memories lose the surrounding context needed to judge whether they provide valid evidence for the current query. To address this problem, we propose Contextual Reinstatement for Agentic Memory (RaMem), a framework that turns retrieved memory fragments into contextually verifiable evidence. RaMem operates through four coordinated stages: (i) evidence anchoring grounds each memory in its original episodic conditions, especially event time, mention time, session span, and participants; (ii) recall condition induction derives the evidence conditions implied by the query; (iii) validity-aware retrieval uses these conditions to prioritize context-compatible memories while retaining content-relevant candidates as fallback evidence; and (iv) context-preserved synthesis keeps the selected memories' structured context available to the generator. Experiments on long-term memory benchmarks show that RaMem consistently improves performance over strong memory baselines, with average F1 gains of more than 10% across several backbones.