acceptodds
Under review as a conference paper at ICLR 2027

Cued Recall: Making Long-Term Memory Accessible with Relationally Conditioned Cues

Abstract

Long-term memory systems for large language model (LLM) agents typically match a query against the event's own content. This practice creates a representation asymmetry between query formulation and memory storage. Anticipated-question generation can narrow this gap when stored information is expressed in query-like forms, but content-only generation remains tied to the event's own wording and may miss questions whose answers depend on its relational context. We propose Cued Recall, a memory framework that grounds anticipated-question generation in local relational structure and aligns memory representations with the diverse ways future queries may access them. Specifically, Cued Recall first constructs a relation-labeled event graph. It links events through semantic, keyword, temporal, participant, and causal relations. Then, for each event, an LLM generates anticipated questions conditioned on its local relational neighborhood. Each question is grounded in information supported by the event and its context. At query time, the cues are combined with event content to identify seed events, and relation-aware graph activation retrieves supporting evidence around those seeds. On LoCoMo, the removal of generated cues causes a 12.20% relative drop in F1. A controlled comparison shows that cue generation grounded in relational context, rather than in the event's content alone, improves F1 by 6.89 points overall and by 10.89 points on multi-hop questions. Cued Recall further achieves an average F1 of 69.37 on HotPotQA and an average accuracy of 62.31% on PersonaMem.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Long-Term Memory;Memory Retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_G3G9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Cued Recall memory framework to address the representational asymmetry between query expression and memory storage form in long-term memory systems of large models. Drawing on the distinction between memory availability and accessibility in cognitive psychology, the framework takes retrieval accessibility as the optimization objective during the memory construction phase: it first constructs an event graph containing five types of relations: semantics, keywords, temporal order, participants, and causality. Then, conditioned on the local relational neighborhood of each event, the LLM generates retrieval cues oriented toward future queries. In the retrieval phase, seed events are identified through three-channel matching of content, keywords and generated cues, followed by relation-aware graph spreading activation to supplement supporting evidence.

Critical strengths
  1. Rather than just tweaking retrieval pipelines, this paper starts from the classic availability-accessibility distinction in human memory and argues that accessibility should be built into memory representations at construction time, not patched in later at query time. It bridges the representation gap between how people ask questions and how information is stored by pre-generating relationally conditioned retrieval cues.
  2. The paper does not stop at reporting final performance numbers. Through fine-grained ablations, controlled comparisons of text-only versus graph-conditioned cues, and gold-standard evidence coverage analysis, it unpacks where the gains actually come from. It clearly separates the contribution of generated cues, the incremental value of relational conditioning, and the complementary role of graph activation.
Critical weaknesses
  1. Generative retrieval representations, spreading activation on graphs, and relational memory are all established ideas. The main novelty here is combining them and conditioning cue generation on an event’s local graph neighborhood. There is no fundamental breakthrough in how cues are generated or how graph traversal works. Essentially this is a thoughtful adaptation and combination of existing techniques to the long-term memory setting.
  2. The paper only reports one-time construction cost on static conversation snapshots. In production, conversational memory updates incrementally turn by turn, but the overhead of extracting entities and relations, updating the graph, and regenerating cues for each new message is never measured. If every update requires a full rebuild of the graph and cue set, costs would scale poorly for long conversations. If incremental updates are used instead, it is not clear how that affects relation consistency or cue quality. This is a central practical concern that is not addressed.
  3. The design retains only the single strongest relation type per event pair, which discards complementary information from multi-dimensional links. Two events might be close in time and semantically related, for example, but only one of those signals survives. Meanwhile, graph activation is capped at 2 hops and never leverages richer topological structure like paths, communities, or node centrality. The graph ends up functioning mostly as an entity association index to expand recall, rather than unlocking deeper structured reasoning like multi-hop inference or causal chain tracing.
Questions for authors
  1. All construction experiments in the paper use static conversation snapshots. In a real continuous conversation setting, how does the system handle new incoming messages? Does each new turn require a full rebuild of the event graph and cues? Is there an incremental update path, and if so what are the performance and cost numbers?
  2. The current design keeps only the strongest relation type per event pair, which can drop information from multiple coexisting associations (e.g., two events sharing both temporal and semantic links). Is there a variant that preserves multiple relation labels? How does performance shift in that setting?
  3. Graph activation is currently limited to a maximum depth of 2 hops. Would deeper propagation (3 hops or more) yield further improvements, or does performance fall off due to accumulated noise? Are there experimental results exploring this?
5 comments

We thank you for taking the time to read our work and for offering constructive suggestions. We appreciate your recognition of the importance of retrieval-accessible long-term memory, as well as your comments on how to better establish the contribution of Cued Recall.

The central contribution of Cued Recall is relationally conditioned memory encoding. During memory construction, the system generates future-oriented retrieval cues from a focal event together with its local, relation-labeled neighborhood. The resulting cues can represent not only the content of an event, but also how that event is connected to other events that may be needed to answer a future question. At retrieval time, the stored cues help identify initial evidence, while the event graph provides a structure for expanding and organizing that evidence. We do not claim to introduce hypothetical-question generation or graph traversal in isolation; our contribution is the way relational context is incorporated into cue construction and carried through to retrieval.

We respond to your comments below by order.

1.1. Contribution beyond combining existing components

"Generative retrieval representations, spreading activation on graphs, and relational memory are all established ideas. The main novelty here is combining them and conditioning cue generation on an event’s local graph neighborhood."

You are right that the individual components of Cued Recall have precedents. We do not claim to introduce a new hypothetical-question generation method, graph-based memory structure, or spreading-activation algorithm. We agree that the contribution lies in combining these ideas, particularly by conditioning cue generation on an event's local relational neighborhood. We would like to clarify what makes this integration different from simply placing existing components in the same pipeline.

In Cued Recall, the event graph does not serve only as a structure for expanding retrieval candidates. Its relational information is used earlier, during memory construction, when the LLM generates future-oriented retrieval cues from a focal event and its relation-labeled neighbors. This allows the generated cues to anticipate access paths that depend on relationships between events, rather than relying solely on the wording of an individual event. The resulting cues are stored as retrieval representations to identify initial event seeds at query time, while the same event graph is subsequently used to expand those seeds with relationally connected evidence. Thus, relational structure connects two stages that would otherwise operate independently: the construction of representations for future access and the retrieval of supporting evidence.

System-level comparison with existing methods. The original LoCoMo main results provide evidence for the effectiveness of the complete system against established memory and retrieval baselines. Cued Recall achieves the highest Overall F1 and Multi-hop F1 among the compared methods on both backbones. Its advantage also extends to single-hop and temporal questions in the main evaluation. Open-domain questions remain more challenging, and Cued Recall does not achieve the best score in every category. These results establish the effectiveness of the complete system, although they do not by themselves identify which design choices account for its gains.

Controlled comparison of relational conditioning. To examine the role of the specific encoding design, we further consider the Event-only, Plain-neighbor, and Full conditions. This is a separate controlled evaluation on the full 1,540-question non-adversarial LoCoMo set, using Qwen3-8B and GPT-4o-mini with three random seeds (42, 123, and 2026), while keeping the other settings fixed. We report the mean and sample standard deviation across the three runs.

QA backboneConditionOverall F1Multi-hop F1
Qwen3-8BEvent-only40.66 ± 0.1129.58 ± 0.11
Qwen3-8BPlain-neighbor42.18 ± 0.1433.26 ± 0.34
Qwen3-8BCued Recall (Full)46.66 ± 0.0738.05 ± 0.48
GPT-4o-miniEvent-only50.78 ± 0.1338.39 ± 0.33
GPT-4o-miniPlain-neighbor52.65 ± 0.1343.40 ± 0.25
GPT-4o-miniCued Recall (Full)57.65 ± 0.1745.46 ± 0.38

Event-only generates cues from the focal event alone, while Plain-neighbor additionally provides neighboring event texts without relation labels or relation-organized output. Full uses relation-labeled neighbors and organizes cue generation by relation dimensions. Adding untyped neighboring text already improves over Event-only, indicating that neighboring context is useful. However, Full provides further gains over Plain-neighbor: 4.48 points in Overall F1 and 4.79 points in Multi-hop F1 with Qwen3-8B, and 5.00 and 2.06 points, respectively, with GPT-4o-mini. This suggests that the improvements cannot be explained by simply providing more event text. They are associated with the joint use of relation labels and relation-organized cue generation.

Evidence ranking at retrieval time. We also examine whether the differences in cue construction are reflected in the initial evidence selected by the retrieval pipeline. The following supplementary analysis is based on the actual fused runtime seed list, before graph activation, and evaluates multi-hop questions against their annotated Gold Evidence nodes. Gold R@K measures the fraction of all annotated Gold Evidence nodes retrieved among the top K runtime seeds. Gold MRR averages reciprocal rank over those nodes, assigning zero to unretrieved nodes.

MethodGold MRRGold R@5Gold R@15
No-Cue0.088414.2%18.8%
Event-only0.142723.1%46.2%
Plain-neighbor0.169525.7%44.2%
Relation-labels0.171329.3%45.0%
QuOTE0.184131.7%43.5%
HyPE0.189927.4%50.4%
Doc2Query0.120315.4%31.7%
Doc2Query + Filter0.118815.4%30.0%
Cued Recall (Full)0.209334.0%50.4%

Full achieves the highest observed Gold MRR and Gold R@5, and ties HyPE at Gold R@15. Compared with Plain-neighbor, Gold MRR increases from 0.1695 to 0.2093, and Gold R@5 from 25.7% to 34.0%. Since this analysis is performed before graph activation, it reflects the quality of the initial runtime seeds rather than the independent contribution of subsequent graph expansion.

Taken together, the external benchmark comparison demonstrates the effectiveness of the complete system, while the controlled comparison and runtime seed analysis provide more specific evidence about the role of relationally conditioned cue construction.

Our contribution is that typed relations among conversational events shape the retrieval representations constructed before a query arrives and are then reused to guide the expansion of retrieved evidence. This is the specific integration that distinguishes Cued Recall from simply combining a generic question generator with an existing graph retriever.

1.2. Incremental memory updates

"In a real continuous conversation setting, how does the system handle new incoming messages? Does each new turn require a full rebuild of the event graph and cues? Is there an incremental update path, and if so what are the performance and cost numbers?"

You raise an important question about maintaining memory as conversations evolve. One relevant design choice in Cued Recall is that generated cues are not rewritten when new memories are added. This differs from A-Mem, which uses an LLM to rewrite neighboring memory representations upon the insertion of a new memory. As also noted in another review, avoiding such rewriting is an intentional choice: once a cue has been generated from an event and its relational context, it remains unchanged rather than triggering repeated LLM-based updates to existing cues whenever the memory grows. This avoids a potentially substantial source of repeated construction cost. Our construction-cost comparison in Table 4 reflects this design choice, showing substantially lower token usage and LLM calls than A-Mem while maintaining competitive retrieval performance.

However, avoiding cue rewriting does not eliminate the need to maintain the event graph as new messages arrive. A new event may introduce additional relations to existing events, and these changes may affect the relational neighborhoods used for future cue generation and graph-based retrieval. Our current experiments construct the event graph and cue representations from static conversation snapshots and cache them for subsequent retrieval. We have not yet implemented or evaluated a complete online update mechanism that incrementally detects affected relations and selectively updates the corresponding graph structures or regenerates cues when necessary.

We therefore distinguish two aspects of the design: existing cues are immutable after construction, avoiding repeated rewriting of previously stored representations; incremental graph maintenance remains future work, particularly for identifying newly affected event relations and keeping the graph consistent as conversations evolve. A practical extension would update the graph locally when new events arrive and regenerate cues only when changes materially affect an event's relational context. Evaluating the cost and retrieval impact of such selective updates is an important direction for future work.

1.3. Single strongest relation versus multiple coexisting relations

"The current design keeps only the strongest relation type per event pair, which can drop information from multiple coexisting associations (e.g., two events sharing both temporal and semantic links). Is there a variant that preserves multiple relation labels? How does performance shift in that setting?"

You asked whether preserving all coexisting relation labels would improve the system. We ran an additional end-to-end comparison between a Single-label condition, which retains only the strongest relation for each event pair, and a Multi-label condition, which preserves all coexisting relation labels and passes them to relation-conditioned cue generation.

Both conditions use the same 12-unique-neighbor cap, fixed 10-seed plus 5-activation visible-event budget, Qwen3-8B QA model, no-tool direct answering, and 15-event budget. We evaluate 1,540 paired LoCoMo questions, excluding Category 5.

Single-label versus Multi-label

ScopeSingle-label F1Multi-label F1Single-label BLEU-1Multi-label BLEU-1
Overall0.46540.46330.40540.4013
Multi-hop0.40630.36950.34090.3026
Temporal0.52750.55650.46780.4923
Open-domain0.24140.23370.12580.1194
Single-hop0.47900.47340.42720.4203

Preserving all relation labels does not improve overall QA in this controlled evaluation. On multi-hop questions, Multi-label is numerically lower on both F1 and BLEU-1.

This result suggests that retaining all coexisting relation labels does not automatically improve downstream cue generation or QA. Additional labels may introduce competing or less useful conditioning signals. Under the evaluated setting, Single-label is slightly higher on overall F1 and BLEU-1, while Multi-label performs better on the temporal subset. These results support retaining the strongest relation as a simpler default for the current system, but do not establish that single-label representations are universally preferable. The conclusion is limited to the variant and evaluation conditions tested here.

Reviewer_YnCyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses long-term conversational memory for LLM agents: given a long multi-session interaction history and a later query, retrieve the relevant past events and answer. The authors argue that existing memory systems match queries against an event's own content, so events reached through a relation (participant, time, cause) rather than their wording are missed. The proposed system, Cued Recall, segments the history into events and builds a relation-labeled graph with five edge types (semantic, keyword, temporal, participant, LLM-judged causal), with ECDF-calibrated scores and one retained relation per pair. For each event, an LLM generates some anticipated questions conditioned on the event and a small sample of its graph neighbours. At query time, they run a two-hop spreading activation with query-type-dependent relation weights to add 20 further events, which are passed to the answer LLM. Validation is on LoCoMo, streaming HotPotQA and PersonaMem and the authors show enouraging performance gains.

Critical strengths
  1. The idea of graph expansion and pre-indexing events into a graph pre-retrieval is very interesting and needs more attention. This work is a nice application of the metadata one can generate and the pre-compute one can do before hand to solve long context conversational QA>

  2. The empirical gains over current agent-memory baselines are consistent across two backbones and three benchmarks. This matters because it shows the combination of index-time question expansion and graph expansion is practically effective for this setting.

Critical weaknesses

The central novelty claim is that conditioning cue generation on relational context, rather than generating cues at all, drives the gain. This claim rests on a single experiment in Table 5. This experiment has no variance or significance reporting. The main results use a different backbone. In that experiment, text-only cues perform slightly worse than no cues overall, and markedly worse on multi-hop questions (35.78 to 30.16). This contradicts a substantial body of evidence that doc2query-style expansion helps retrieval, which suggests the text-only baseline is weak or its cues are noisy. I was looking for standard docT5query-style baseline or hallucination filtering (as in Doc2Query--) is reported. Without a full-scale, multi-backbone, multi-seed version of this comparison, the paper's contribution reduces to an engineering combination of known components.

  1. The positioning and conceptual contribution are overstated relative to prior information retrival work. The availability/accessibility framing relabels vocabulary mismatch, and index-time generation of anticipated queries is essentially doc2query. Conditioning expansion on surrounding context has precedents, such as contextual chunk augmentation and contextual document embeddings. The retrieval stage is seed-then-expand over a corpus graph, which is the setting studied in the graph-based adaptive retrieval line (GAR, LADR, and recent follow-up work on adaptive retrieval like Slidegar, and QUAM). None of this is cited or compared against. the paper's own ablation shows the expansion policy matters (uniform expansion is worse than no expansion) so these baselines become meaningful comparisons. At least the authors should argue or position against them.

  2. Some other soundness issues are lack of expected controls — a) Up to 60 events reach the answer LLM, with no control for context budget against baselines. b) Hyperparameters are swept on the evaluation benchmark whi has the potential of leakage or overfitting.

  3. Final observation that I was not able to reconcile — the cue-only seeds cover only 25% of gold evidence despite much higher cue-only R@10; this suggests cues may route queries to neighbouring events, i.e., multi-hop joins precomputed at index time, and that is not analyzed.

Questions for authors
  1. Can you report the graph-conditioned vs. text-only vs. no-cue comparison on the full LoCoMo set, with both backbones, multiple seeds and significance tests?

  2. Can you add a standard doc2query baseline, with and without filtering of unsupported questions?

  3. Why does text-only expansion hurt multi-hop performance?

4 comments

We thank you for taking the time to read our work and for offering constructive suggestions. We appreciate your recognition of the importance of retrieval-accessible long-term memory, as well as your comments on how to better establish the contribution of Cued Recall.

The central contribution of Cued Recall is relationally conditioned memory encoding. During memory construction, the system generates future-oriented retrieval cues from a focal event together with its local, relation-labeled neighborhood. The resulting cues can represent not only the content of an event, but also how that event is connected to other events that may be needed to answer a future question. At retrieval time, the stored cues help identify initial evidence, while the event graph provides a structure for expanding and organizing that evidence. We do not claim to introduce hypothetical-question generation or graph traversal in isolation; our contribution is the way relational context is incorporated into cue construction and carried through to retrieval.

We respond to your comments below by order.

1.1. Full-dataset evaluation across two backbones and three random seeds

"Can you report the graph-conditioned vs. text-only vs. no-cue comparison on the full LoCoMo set, with both backbones, multiple seeds and significance tests?"

You asked for the comparison on the full LoCoMo set, using multiple QA backbones and random seeds. We have replaced the earlier small-subset comparison in Section 5.4 with an evaluation on the full 1,540-question non-adversarial LoCoMo set, using Qwen3-8B and GPT-4o-mini.

We ran three independent seeds (42, 123, and 2026). The random seed was the only setting varied across these runs; the prompt, temperature, top-p, retrieval budget, and other experimental settings were held fixed. We report the mean and sample standard deviation across the three runs. These repeated runs allow us to assess whether the observed differences are stable across random initializations. To avoid ambiguity, the former Text-only condition is now called Event-only, since it generates cues from the focal event alone.

Qwen3-8B (mean ± sample standard deviation, %)

MethodOverallSingle-hopMulti-hopTemporalOpen-domain
No-Cue37.30 ± 0.0543.06 ± 0.0726.26 ± 0.1738.97 ± 0.3513.74 ± 0.25
Event-only40.66 ± 0.1147.42 ± 0.1929.58 ± 0.1139.45 ± 0.1918.08 ± 0.13
Plain-neighbor42.18 ± 0.1448.52 ± 0.2533.26 ± 0.3440.51 ± 0.1318.41 ± 0.24
QuOTE42.89 ± 0.1651.17 ± 0.2131.08 ± 0.3739.66 ± 0.1415.94 ± 0.13
HyPE41.81 ± 0.1548.25 ± 0.1933.79 ± 0.2739.29 ± 0.2317.33 ± 0.35
Doc2Query40.73 ± 0.0448.51 ± 0.1029.50 ± 0.3536.96 ± 0.1618.12 ± 0.37
Doc2Query + Filter40.84 ± 0.3347.83 ± 0.5329.56 ± 0.2539.24 ± 0.1618.12 ± 0.37
Cued Recall (Full)46.66 ± 0.0752.90 ± 0.1438.05 ± 0.4846.38 ± 0.1718.16 ± 0.11

GPT-4o-mini (mean ± sample standard deviation, %)

MethodOverallSingle-hopMulti-hopTemporalOpen-domain
No-Cue47.41 ± 0.2651.82 ± 0.3436.27 ± 0.2153.85 ± 0.2819.91 ± 0.10
Event-only50.78 ± 0.1356.13 ± 0.1638.39 ± 0.3355.01 ± 0.0826.15 ± 0.14
Plain-neighbor52.65 ± 0.1357.29 ± 0.3743.40 ± 0.2556.49 ± 0.3226.32 ± 0.14
QuOTE53.38 ± 0.0859.35 ± 0.2942.18 ± 0.2456.57 ± 0.5123.33 ± 0.38
HyPE51.95 ± 0.1556.50 ± 0.3344.02 ± 0.1555.03 ± 0.1225.10 ± 0.28
Doc2Query50.90 ± 0.1257.16 ± 0.2638.33 ± 0.2152.94 ± 0.1826.26 ± 0.20
Doc2Query + Filter51.38 ± 0.1457.36 ± 0.2538.68 ± 0.2254.37 ± 0.1226.24 ± 0.14
Cued Recall (Full)57.65 ± 0.1763.53 ± 0.3845.46 ± 0.3862.26 ± 0.2126.52 ± 0.25

Full achieves the highest overall and multi-hop F1 among the evaluated cue-generation methods for both backbones. Compared with Plain-neighbor, Full improves overall F1 by 4.48 points and multi-hop F1 by 4.79 points with Qwen3-8B. With GPT-4o-mini, the corresponding improvements are 5.00 and 2.06 points. Compared with Event-only, Full improves multi-hop F1 by 8.47 points with Qwen3-8B and 7.07 points with GPT-4o-mini.

The standard deviations are small across the three tested seeds, indicating that the reported differences are stable under these runs.

1.2. Doc2Query and filtering unsupported questions

"Can you add a standard doc2query baseline, with and without filtering of unsupported questions?"

You asked us to include Doc2Query both with and without filtering questions that cannot be answered from their source text. We added both controlled variants to the full-dataset comparison.

Doc2Query generates up to four natural questions from each focal event using a Doc2Query-style prompt. We concatenate the original event text and the generated questions into one BM25 document:

Event: [event text] Hypothetical questions: [q1, q2, q3, q4]

The filtered variant adds an LLM-based check to determine whether each generated question can be answered from its source event alone. Questions judged unsupported by that event are discarded. Importantly, this filter checks source-event answerability only; it does not inspect neighboring events or verify cross-event relational consistency.

In the full-dataset results, filtering yields small numerical gains across all four reported comparisons. With Qwen3-8B, overall F1 increases by 0.11 points and multi-hop F1 by 0.06 points. With GPT-4o-mini, the corresponding gains are 0.48 and 0.35 points. These improvements are modest and do not close the gap to Cued Recall, suggesting that source-event answerability filtering alone is insufficient to address the cross-event retrieval challenge targeted by our method.

1.3. Why did the previous Text-only condition appear to hurt multi-hop performance?

"Why does text-only expansion hurt multi-hop performance?"

You asked about the multi-hop decrease reported for the former Text-only condition. That result in the original Table 5 was obtained from a randomly sampled, much smaller subset of LoCoMo. The full-dataset, three-seed evaluation does not reproduce the decrease.

The former Text-only condition, now named Event-only, improves over No-Cue on multi-hop questions with both backbones:

BackboneNo-Cue multi-hop F1Event-only multi-hop F1Difference
Qwen3-8B26.2629.58+3.32 pp
GPT-4o-mini36.2738.39+2.12 pp

The earlier result was therefore not representative of the full-dataset behavior and appears to have been sensitive to the sampled subset. We have replaced the small-subset result with the full-dataset evaluation and clarified the terminology in the revised manuscript.

1.4. Actual seed ranking and Gold Evidence coverage

"The cue-only seeds cover only 25% of gold evidence despite much higher cue-only R@10; this suggests cues may route queries to neighbouring events, i.e., multi-hop joins precomputed at index time, and that is not analyzed."

You noticed an apparent discrepancy between the original cue-only R@10 and Gold Evidence Coverage. These metrics use different denominators. The original retriever-level R@10 is a query-level hit rate: a question counts as a hit if at least one annotated Gold Evidence node appears in the top 10. Gold Evidence Coverage is a node-level measure: it counts the fraction of all annotated gold evidence nodes retrieved.

For example, if a question has four annotated Gold Evidence nodes and the retrieval result contains one of them, it counts as a query-level hit, while its node-level coverage is only 25%. A high query-level R@10 can therefore coexist with much lower node-level coverage.

To examine the actual runtime behavior, we additionally analyzed the fused seed list selected by the retrieval pipeline before graph activation. This is a system-level analysis of the seeds actually passed to activation, not an offline ranking of content and cue candidates separately. We use the same Gold Evidence annotations and evaluate the multi-hop questions. The reported R@K is node-level recall: for each question, it measures the fraction of annotated Gold Evidence nodes appearing in the top K runtime seeds. Formally, for question q with Gold Evidence set G_q:

= \frac{|G_q \cap \mathrm{TopK}(q)|}{|G_q|}.$$ Gold MRR averages the reciprocal rank over the annotated Gold Evidence nodes, assigning zero to a node that is not retrieved. These metrics therefore evaluate how well the actual fused seed list ranks all required evidence nodes, rather than whether it retrieves at least one gold node for a question. **Actual runtime seed ranking on multi-hop questions** | Method | Gold MRR | Gold R@5 | Gold R@15 | | --- | --- | --- | --- | | No-Cue | 0.0884 | 14.2% | 18.8% | | Event-only | 0.1427 | 23.1% | 46.2% | | Plain-neighbor | 0.1695 | 25.7% | 44.2% | | Relation-labels | 0.1713 | 29.3% | 45.0% | | QuOTE | 0.1841 | 31.7% | 43.5% | | HyPE | 0.1899 | 27.4% | 50.4% | | Doc2Query | 0.1203 | 15.4% | 31.7% | | Doc2Query + Filter | 0.1188 | 15.4% | 30.0% | | **Cued Recall (Full)** | **0.2093** | **34.0%** | **50.4%** | Full achieves the highest observed Gold MRR and Gold R@5, and ties HyPE at Gold R@15. The results indicate that, in the actual runtime seed list, Full ranks annotated Gold Evidence more favorably at the early retrieval stage. The advantage is clearest at a small seed budget (R@5), while the R@15 result is on par with HyPE.

1.5. Positioning against contextual augmentation and graph-based adaptive retrieval

"The availability/accessibility framing relabels vocabulary mismatch ... Conditioning expansion on surrounding context has precedents, such as contextual chunk augmentation and contextual document embeddings. The retrieval stage is seed-then-expand over a corpus graph, which is the setting studied in the graph-based adaptive retrieval line (GAR, LADR, and recent follow-up work on adaptive retrieval like Slidegar, and QUAM)."

You raise two related points: contextual information has been used to enrich retrieval representations, and graph-based adaptive retrieval has explored seed-then-expand strategies. We agree that neither contextual augmentation nor retrieval-time graph expansion is new in itself.

Contextual chunk augmentation, such as Contextual Retrieval, adds surrounding document context to individual chunks before indexing. Cued Recall also uses context during memory construction, but the context is different: it consists of typed relations among conversational events, including temporal, participant, semantic, keyword, and causal relations. These relations provide explicit conditioning signals for generating future-oriented cues, rather than simply adding descriptive context to the focal event's text or embedding.

GAR, LADR, SlideGAR, and QuAM use corpus or document relationships to guide candidate expansion or reranking at retrieval time. Their graphs may be constructed offline; our distinction is not that graph construction happens at a different time. Rather, Cued Recall uses the same typed event neighborhood at two stages: before a query arrives, it conditions cue generation during memory construction; after a query arrives, the event graph is reused to expand the initial seeds with relationally connected evidence. The contribution is using typed conversational relations to shape stored retrieval representations before a query arrives and to support evidence expansion afterward.

We have also clarified that the availability/accessibility distinction motivates this design rather than introducing vocabulary mismatch as a new problem. Our empirical claim is narrower: in the evaluated conversational-memory setting, relationally conditioned cue construction improves over event-only cue generation and untyped neighboring context, with corresponding gains in downstream QA and initial Gold Evidence ranking.

We hope these explanations and additional experiments address your initial concerns. Thank you once again for your suggestions; we look forward to receiving your further feedback.

Reviewer_VqrbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

One-line Summary.

This paper proposes Cued Recall, a memory framework for long-term LLM agents that leverages anticipated future questions for each stored event. These anticipated questions are generated by constructing the event's relation-labeled graph and using the graph's neighborhood.

 

Limitation of Prior Work.

The authors explain two limitations of existing long-term memory systems:

  • Accessibility is considered only in the retrieval phase, not in the construction phase.
  • Use of content-only generated representations, rather than also considering other relations such as semantic, temporal, and causal relationships between events.

 

Core Idea.

  • Relation-labeled graph construction: events are linked through five relation types (semantic, keyword, temporal, participant, causal).
  • Cue construction: each event stores three cue channels, C_verb (keywords and entities), C_sem (event text and paraphrases), and C_gen (future-oriented questions generated by an LLM from the event together with a bounded relational neighborhood).

 

Effect.

Cued Recall shows the best or second-best performance over existing baselines when evaluated on LoCoMo, HotPotQA, and PersonaMem.

Critical strengths

(1) Clear retrieval benefit of the Content Cue (Table 3, Table 6(a))

Table 3 and Table 6(a) support the benefit of the Content Cue, which greatly advances retrieval performance. This is a good realization of the authors' claim that accessibility should be constructed at construction time rather than left to query-time search.

 

(2) A good strategy not to rewrite cues after construction

While the basic idea seems to be similar to A-Mem, which also uses the concept of neighboring notes (e.g., keywords, tags, and descriptions) to improve retrieval, I agree with the authors' design choice to give up "rewriting". A-Mem rewrites the representations of neighboring notes with an LLM whenever a new memory is added, whereas Cued Recall does not rewrite a cue once it is generated. Table 4 shows that this design leads to a large reduction in construction cost compared with A-Mem while achieving comparable performance.

 

(3) Detailed hyperparameter search

Figure 2 provides one-factor sweeps over six hyperparameters, giving users a helpful overview of how the framework behaves.

Critical weaknesses

(1) Unclear novelty against the closest line of prior work

One of my concerns after reading this paper was novelty. In my view, the authors' key idea consists of two parts: (a) anticipated-question generation, and (b) achieving this through a graph construction with relational conditioning (e.g., semantic/temporal/etc.).

However, (a) is not a new idea; the authors acknowledge this, but represent this idea only with older methods such as doc2query, Query2doc, and HyDE. There are more recent LLM-based methods such as QuOTE and HyPE, which generate multiple anticipated questions per chunk with an LLM and index them for retrieval. The authors should include this line of work in the related work section, clarifying the differences against these works to support novelty.

References.
Neeser et al., QuOTE: Question-Oriented Text Embeddings, arXiv 2025.
Vake et al., Bridging the Question-Answer Gap in Retrieval-Augmented Generation: Hypothetical Prompt Embeddings (HyPE), IEEE 2025.

 

(2) Lack of experimental support for justifying why graph-based relational conditioning is important

Continuing from (1), it seems to be important to support why (b) is important (because this differentiates this paper from other recent LLM-based work). However, the current experiments do not isolate it from simply adding more context to the generator.

More specifically, (b) can be broken down into three parts:

  • (i) the text of neighboring events is added to the generator input
  • (ii) the relation label of each neighbor (one of the five types) is provided
  • (iii) the output is organized by relation dimension (at most 3 cues per dimension).

The current comparison is between a condition with none of (i)-(iii) (Text-only Cue) and one with all of them (Graph-conditioned Cue). To support the effect of graph conditioning, the authors should add a condition with (i) only, i.e., the same neighbors given as plain text without relation labels or dimension structure.

 

(3) Missing strongest baselines in experiments

  • Missing E-mem in LoCoMo (Table 1): E-mem is compared only on HotPotQA and is marked as "HotPotQA only", while the original E-mem paper reports LoCoMo results. Since E-mem is one of the strongest baselines (Table 2), I encourage the authors to explain why it was excluded.
  • Missing SimpleMem and E-mem in construction cost comparison (Table 4): Similarly, SimpleMem is the strongest baseline in Table 1, but it is excluded from the cost comparison.

 

(4) Missing description of baselines

While I am a bit familiar with the baselines used in the experiments, it is commonly encouraged to include a brief description of each baseline and the rationale for selecting them, yet omitted in current manuscript.

 

Overall, I believe the strategy is reasonable, and the strengths above are real, but my score currently reflects (1) and (2), which concern the paper's central claim about relational conditioning.

Questions for authors

Q. Readability of Figure 2

In Figure 2, the subplot titles overlap with the plots and the axis labels, which makes the figure hard to read. Could the authors adjust the layout for better readability?

5 comments

We thank you for taking the time to read our work and for offering constructive suggestions. We appreciate your recognition of the importance of retrieval-accessible long-term memory, as well as your comments on how to better establish the contribution of Cued Recall.

The central contribution of Cued Recall is relationally conditioned memory encoding. During memory construction, the system generates future-oriented retrieval cues from a focal event together with its local, relation-labeled neighborhood. The resulting cues can represent not only the content of an event, but also how that event is connected to other events that may be needed to answer a future question. At retrieval time, the stored cues help identify initial evidence, while the event graph provides a structure for expanding and organizing that evidence. We do not claim to introduce hypothetical-question generation or graph traversal in isolation; our contribution is the way relational context is incorporated into cue construction and carried through to retrieval.

We respond to your comments below by order.

1.1. Novelty relative to QuOTE and HyPE

"There are more recent LLM-based methods such as QuOTE and HyPE, which generate multiple anticipated questions per chunk with an LLM and index them for retrieval. The authors should include this line of work in the related work section, clarifying the differences against these works to support novelty."

You are right that generating hypothetical questions for stored text is an established way to improve retrieval. We have therefore added QuOTE and HyPE to the Related Work section of the revised manuscript (highlighted in blue). We have also clarified that Cued Recall does not claim novelty in generating anticipated questions itself.

QuOTE and HyPE generate hypothetical questions or prompts for individual text chunks and use these generated texts to enrich the retrieval representation of each chunk. Cued Recall addresses a different aspect of cue construction: the LLM receives not only the focal event, but also its local, relation-labeled event neighborhood. The generated cues can therefore encode how an event is connected to other events, rather than describing only the focal event in isolation.

This distinction matters in conversational memory because a future question may refer to an event indirectly—for example, through its temporal, participant, semantic, keyword, or causal connection to another event. The relevant access path may not be explicit in the wording of the focal event itself. By conditioning cue generation on these relations during memory construction, Cued Recall stores representations that anticipate such cross-event access patterns before the future query is known. This does not mean that chunk-level question generation can never infer cross-event associations; rather, Cued Recall explicitly provides typed inter-event relations as conditioning context instead of relying on the generator to infer them from the focal event alone.

We have also included controlled implementations of QuOTE and HyPE in our full-dataset comparison. To make the comparison clear without requiring you to consult another section(In section 5.4), the overall and multi-hop F1 results are summarized below. All values are means over three random seeds.

QA backboneMethodOverall F1Multi-hop F1
Qwen3-8BQuOTE42.89 ± 0.1631.08 ± 0.37
Qwen3-8BHyPE41.81 ± 0.1533.79 ± 0.27
Qwen3-8BCued Recall (Full)46.66 ± 0.0738.05 ± 0.48
GPT-4o-miniQuOTE53.38 ± 0.0842.18 ± 0.24
GPT-4o-miniHyPE51.95 ± 0.1544.02 ± 0.15
GPT-4o-miniCued Recall (Full)57.65 ± 0.1745.46 ± 0.38

Across both QA backbones, Cued Recall achieves higher overall and multi-hop F1 than the controlled QuOTE and HyPE implementations. We report these results as evidence for the value of relational conditioning, not as evidence that hypothetical-question generation itself is new.

Our availability/accessibility framing motivates this design rather than serving as a standalone novelty claim. The technical contribution is the use of relational context to shape the representations stored for future access: the neighborhood conditions cue generation before a query arrives, and the same event structure is reused to expand evidence at retrieval time.

1.2. Isolating the contribution of relational conditioning

"The current comparison is between a condition with none of (i)-(iii) (Text-only Cue) and one with all of them (Graph-conditioned Cue). To support the effect of graph conditioning, the authors should add a condition with (i) only, i.e., the same neighbors given as plain text without relation labels or dimension structure."

You asked for a control that provides neighboring events without their relation labels or dimension-organized structure. We have added this Plain-neighbor condition to Section 5.4 of the revised manuscript. We have also renamed the previous Text-only condition to Event-only, which more accurately describes its input.

The three conditions are defined as follows:

MethodGeneration inputCue-generation formatIndexed representationRetrieval
Event-onlyFocal event onlyNo relation labels or dimension-organized outputGenerated cue representationsDense
Plain-neighborFocal event + untyped neighbor textsNo relation labels or dimension-organized outputGenerated cue representationsDense
FullFocal event + relation-labeled local neighborhoodRelation-aware, dimension-organized cuesGenerated cue representationsDense

Event-only and Plain-neighbor use the same cue-generation format; the difference is whether untyped neighboring event texts are provided. Full additionally supplies relation labels and organizes cue generation by relation dimensions.

On the full 1,540-question non-adversarial LoCoMo set, Full improves overall F1 over Plain-neighbor by 4.48 percentage points with Qwen3-8B and 5.00 points with GPT-4o-mini. The corresponding multi-hop improvements are 4.79 and 2.06 points. Full also improves multi-hop F1 over Event-only by 8.47 points with Qwen3-8B and 7.07 points with GPT-4o-mini. For completeness, the full QA results for the three core conditions are listed below.

Qwen3-8B (mean ± sample standard deviation, %)

MethodOverallSingle-hopMulti-hopTemporalOpen-domain
No-Cue37.30 ± 0.0543.06 ± 0.0726.26 ± 0.1738.97 ± 0.3513.74 ± 0.25
Event-only40.66 ± 0.1147.42 ± 0.1929.58 ± 0.1139.45 ± 0.1918.08 ± 0.13
Plain-neighbor42.18 ± 0.1448.52 ± 0.2533.26 ± 0.3440.51 ± 0.1318.41 ± 0.24
Cued Recall (Full)46.66 ± 0.0752.90 ± 0.1438.05 ± 0.4846.38 ± 0.1718.16 ± 0.11

GPT-4o-mini (mean ± sample standard deviation, %)

MethodOverallSingle-hopMulti-hopTemporalOpen-domain
No-Cue47.41 ± 0.2651.82 ± 0.3436.27 ± 0.2153.85 ± 0.2819.91 ± 0.10
Event-only50.78 ± 0.1356.13 ± 0.1638.39 ± 0.3355.01 ± 0.0826.15 ± 0.14
Plain-neighbor52.65 ± 0.1357.29 ± 0.3743.40 ± 0.2556.49 ± 0.3226.32 ± 0.14
Cued Recall (Full)57.65 ± 0.1763.53 ± 0.3845.46 ± 0.3862.26 ± 0.2126.52 ± 0.25

We also conducted an additional analysis of the actual runtime seed ranking, which is not presented as a full table in the revised manuscript. The analysis uses the fused seed list selected by the retrieval pipeline before graph activation, rather than independently ranking content and cue candidates offline. It is evaluated on multi-hop questions using the same Gold Evidence annotations. Gold R@K is node-level recall: for each question, it measures the fraction of annotated Gold Evidence nodes appearing in the top K runtime seeds. Gold MRR averages the reciprocal rank over the annotated Gold Evidence nodes, assigning zero to nodes that are not retrieved.

MethodGold MRRGold R@5Gold R@15
No-Cue0.088414.2%18.8%
Event-only0.142723.1%46.2%
Plain-neighbor0.169525.7%44.2%
Relation-labels0.171329.3%45.0%
QuOTE0.184131.7%43.5%
HyPE0.189927.4%50.4%
Doc2Query0.120315.4%31.7%
Doc2Query + Filter0.118815.4%30.0%
Cued Recall (Full)0.209334.0%50.4%

Full has higher observed values than Plain-neighbor on all three seed-ranking metrics. Together with the QA results, this shows that simply adding untyped neighboring text does not reproduce the results of the full design. The additional gain is associated with the use of relation labels and relation-organized cue generation.

1.3. Why E-mem was initially excluded from the controlled comparison

"Missing E-mem in LoCoMo (Table 1): E-mem is compared only on HotPotQA and is marked as 'HotPotQA only', while the original E-mem paper reports LoCoMo results. Since E-mem is one of the strongest baselines (Table 2), I encourage the authors to explain why it was excluded."

You pointed out that E-mem reports LoCoMo results in its original paper, and we agree that this reference is relevant. We initially did not include E-mem as a controlled LoCoMo baseline because its architecture differs substantially from the single-memory-system implementations in our comparison. E-mem uses a heterogeneous multi-agent architecture in which a GPT-4o-mini master agent coordinates Qwen3-4B assistant agents. These assistants maintain episodic contexts, perform local reasoning, and return evidence for aggregation by the master agent.

We attempted to reproduce E-mem, but could not obtain a faithful reproduction of its complete multi-agent pipeline under our experimental setup. A controlled reproduction would require implementing and coordinating its assistant-agent memory units, routing, local reasoning, and master-agent aggregation. We did not want to present an incomplete implementation as a directly comparable result.

To address your concern, we report the GPT-4o-mini LoCoMo results from the original E-mem paper below as a published cross-paper reference. The Cued Recall values are our full-dataset, three-seed results:

MethodOverall F1Single-hop F1Multi-hop F1Temporal F1Open-domain F1
E-mem (published)54.1759.2342.6459.8224.89
Cued Recall (Full, GPT-4o-mini)57.65 ± 0.1763.53 ± 0.3845.46 ± 0.3862.26 ± 0.2126.52 ± 0.25
Difference in point estimates+3.48+4.30+2.82+2.44+1.63

The original E-mem paper also reports BLEU-1 scores of 44.34 overall, 50.58 for single-hop, 34.38 for multi-hop, 44.57 for temporal, and 18.15 for open-domain questions.

This comparison is informative, but it is not a same-configuration reproduction or a controlled head-to-head test. E-mem uses GPT-4o-mini as the master model together with Qwen3-4B assistant agents, whereas our reported GPT-4o-mini result uses the QA-backbone configuration in our evaluation. We therefore present the values as a reference and do not interpret the point differences as a controlled estimate of the effect of the memory method alone.

1.4. SimpleMem and E-mem in the construction-cost comparison

"Missing SimpleMem and E-mem in construction cost comparison (Table 4): Similarly, SimpleMem is the strongest baseline in Table 1, but it is excluded from the cost comparison."

You also raised the omission of SimpleMem from the construction-cost comparison. We ran an additional measurement for SimpleMem and report it here to address your question. The accounting covers one-time memory construction and includes input tokens, output tokens, total tokens, and LLM calls. We provide this supplementary comparison in the rebuttal rather than adding another table to the revised manuscript.

MethodInput tokens (M)Output tokens (M)Total tokens (M)LLM calls
A-Mem9.1262.36811.49411,754
Mem010.9581.23912.1979,181
MEMORYOS1.8890.9392.8285,534
SimpleMem1.3561.1702.5261,149
Cued Recall2.5441.2023.7461,829

SimpleMem has the lowest measured construction cost in this comparison. Cued Recall is not the cheapest method: its total token use is higher than that of SimpleMem and MEMORYOS, but lower than that of A-Mem and Mem0. In terms of LLM calls, Cued Recall uses more calls than SimpleMem but fewer than MEMORYOS, A-Mem, and Mem0. The results therefore show a mixed cost profile rather than a minimum-cost advantage for Cued Recall.Our approach perhaps achieves a reasonable balance between effectiveness and cost.

We do not include E-mem in this table because its original paper reports a different quantity—per-query inference cost—rather than one-time memory-construction cost. Combining the two would mix different accounting boundaries. We would need a controlled reproduction with the same construction-cost accounting to make a meaningful comparison.

1.5. Baseline descriptions and rationale for selection

"While I am a bit familiar with the baselines used in the experiments, it is commonly encouraged to include a brief description of each baseline and the rationale for selecting them, yet omitted in current manuscript."

You suggested adding a brief explanation of the external baselines and why they were selected. We have added a concise description to the Experimental Setup section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.