acceptodds
Under review as a conference paper at ICLR 2027

DuCoMem: Efficient Long-Term Memory for LLMs through Dual Compression and Local Recovery

Abstract

External memory enables large language models (LLMs) to use long-term interaction histories, but its construction and retrieval costs grow rapidly as histories accumulate. Existing lightweight memory methods, however, compress the memory content to preserve its semantics, overlooking that the details discarded by compression, e.g., names, locations, or temporal cues, are often essential for locating the memory in response to future queries. Meanwhile, graph-based memories rely on the LLMs to construct semantic relations before any query arrives, incurring additional costs. In this paper, we propose DuCoMem, a dual compression memory framework that jointly compresses the memory content and graph structure while preserving on-demand access to the original records. Specifically, we first reveal that the content retention and content retrievability diverge under the compression. The residual addresses are then introduced to store the key details discarded by the compressor, such that the compact memories remain locatable. For graph compression, we first simplify a low-cost candidate graph by removing the redundant nodes and edges, and then use an LLM to check relations only on the smaller graph. During retrieval, DuCoMem starts from a few representative nodes on the compressed graph, reads the nearby original records, and searches further only when the evidence is not enough to answer the question. This design reduces the structure processed during construction and the search space explored during retrieval. Experimental results demonstrate that DuCoMem outperforms the strongest lightweight-memory baseline across model scales, with an average relative improvement of 23.7% in macro F1 across seven answering models on LoCoMo while maintaining a comparable token cost.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, long-term memory, memory-augmented generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_SJXPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents DuCoMem, a long-term memory framework for LLMs that combines content compression, graph coarsening, and on-demand access to original records. Its main observation is that compression may preserve the content needed to answer a question while removing the details needed to find it. DuCoMem keeps these omitted cues as residual addresses, reduces the memory graph while retaining source mappings, and searches for additional evidence when the initial retrieval is insufficient. Experiments on LoCoMo and LongMemEval-S show improvements over the evaluated baselines across multiple answering models, with comparable reported retrieval token costs.

Critical strengths
  1. The paper is clearly written, and the examples make the method easy to follow. The distinction between preserving content and preserving retrievability is useful and well explained. Residual addressing is a straightforward response to this problem, and it fits well with graph coarsening and local source recovery.

  2. The experiment is solid. DuCoMem improves over SimpleMem on LoCoMo across all seven answering models and also performs well on LongMemEval-S. The ablations and compression analysis help explain where these gains come from.

Critical weaknesses
  1. The comparisons cover several established methods, but a few relevant memory systems are missing, such as EverMemOS, SYNAPSE, and Zep. Adding a comparison would give a fuller picture of how DuCoMem performs against other methods.

  2. The ablations could better isolate individual contributions. Removing only residual addressing while keeping the compressed representations fixed, would make the benefits of these components clearer.

Questions for authors
  1. Could the authors add a comparison with EverMemOS, SYNAPSE, or Zep on a shared benchmark?

  2. Could the authors include an ablation that removes only residual addressing?

1 comment

We thank the reviewer for the positive assessment. We summarize our responses below.

  1. On LoCoMo with GPT-4.1-mini and the same scoring script, DuCoMem outperforms SYNAPSE in all four categories, improving macro F1 by 3.98 points (49.72 vs. 45.74) with 43% fewer tokens per question (502 vs. 880).
  2. Removing only residual addressing, with capsules and graph fixed, lowers macro F1 from 49.72 to 48.53.
  3. The largest drop is on SingleHop (62.11 → 56.52).

W1 / Q1: Comparison with SYNAPSE

[W1] The comparisons cover several established methods, but a few relevant memory systems are missing, such as EverMemOS, SYNAPSE, and Zep. Adding a comparison would give a fuller picture of how DuCoMem performs against other methods.

[Q1] Could the authors add a comparison with EverMemOS, SYNAPSE, or Zep on a shared benchmark?

Response:

On LoCoMo with GPT-4.1-mini, DuCoMem outperforms SYNAPSE [C1] in all four question categories, improving Macro F1 by 3.98 points (49.72 vs. 45.74) while using about 43.0% fewer tokens per question (502 vs. 880). We evaluated SYNAPSE on LoCoMo using GPT-4.1-mini and the same scoring script as DuCoMem.

Answer quality and token cost. DuCoMem achieves a Macro F1 of 49.72, 3.98 points higher than SYNAPSE's 45.74, with gains across all four categories. The largest gain is on OpenDomain, where F1 rises from 28.85 to 35.63 (+6.78 points), followed by Temporal, where F1 rises from 56.52 to 60.84 (+4.32 points). Together, these two categories account for about 70% of the Macro F1 gain. DuCoMem also improves MultiHop and SingleHop F1 by 2.96 and 1.85 points, respectively. Under the manuscript's token cost metric, DuCoMem uses 502 tokens per question, about 43.0% fewer than SYNAPSE's 880.

Table C1. SYNAPSE and DuCoMem on LoCoMo with GPT-4.1-mini.

MetricSYNAPSEDuCoMem
MultiHop F137.3240.28
Temporal F156.5260.84
OpenDomain F128.8535.63
SingleHop F160.2662.11
Macro F145.7449.72
Token cost880502

[C1] Jiang et al. Synapse: Empowering LLM Agents with Episodic-Semantic Memory via Spreading Activation. Findings of ACL 2026.

W2 / Q2: Ablation of residual addressing alone

[W2] The ablations could better isolate individual contributions. Removing only residual addressing while keeping the compressed representations fixed, would make the benefits of these components clearer.

[Q2] Could the authors include an ablation that removes only residual addressing?

Response:

Removing only residual addressing, with the compressed capsules, graph structure, and all other retrieval settings fixed, lowers Macro F1 on LoCoMo from 49.72 to 48.53, with the largest drop on SingleHop (62.11 to 56.52). These questions can often be answered from a single record, which residual cues help locate when the compressed capsule omits a distinctive name, place, or date. With residual addressing enabled, DuCoMem also makes fewer follow-up answering calls (94 vs. 103) and uses slightly fewer online tokens per question (502 vs. 505).

Results by question category. As shown in Table C2, the largest drop was on SingleHop, where F1 fell from 62.11 to 56.52. A single record can suffice to answer these questions. When the compressed capsule omits a distinctive name, place, or date, a residual term can still help retrieve that record. MultiHop and Temporal also declined by 0.72 and 0.39 points, respectively. OpenDomain improved by 1.97 points, showing that the effect varies by question category.

Online token cost. With residual addressing disabled, the number of answering calls after follow-up retrieval increased from 94 to 103. These additional calls help explain why total online LLM input and output tokens increased from 502 to 505 per question.

Table C2. LoCoMo F1 and online query cost with and without residual addressing.

MetricFull modelWithout residual addressing
MultiHop F140.2839.56
Temporal F160.8460.45
OpenDomain F135.6337.60
SingleHop F162.1156.52
Macro F149.7248.53
Follow-up answering calls94103
Total online tokens per question502505
Reviewer_jfxpOpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DuCoMem builds long-term conversational memory through content compression and graph coarsening. It retains omitted keywords as retrieval cues and uses source mappings to read original records when needed, with focused search when the initial evidence is insufficient. The authors evaluate it on LoCoMo and LongMemEval-S, reporting improved answer quality at token costs comparable to lightweight memory methods. Residual addressing offers some novelty, and the coarsening results are encouraging. However, the current comparisons do not show what the graph adds over direct retrieval with iterative follow-up search, and the full cost remains unclear.

Critical strengths
  1. Preserving answer content after compression does not guarantee that the memory can still be retrieved. Selecting retrieval cues specifically to complement the compressed representation is a concrete and useful design.
  2. The evaluation covers two datasets and answering models of different sizes. The retrieval diagnostics and component ablations provide some positive evidence, particularly the improvements in both reported accuracy and token cost from coarsening.
Critical weaknesses
  1. Role of the graph. I find the motivation for residual addressing reasonable, but I am less clear about what the graph adds. The collapse rules in Section 3.3 draw on STPGC. ReadAgent already uses summaries to locate and read original text, and GraphReader searches a graph and continues when the evidence is insufficient. The paper needs to explain more directly what coarsening and recovery mappings improve over these existing approaches.

    Table 5 shows that coarsening helps: macro F1 increases from 48.78 to 49.72, while Token Cost drops from 598 to 502. However, w/o Graph Coarsening still uses the original graph and follow-up search, and w/o Focused Search also retains the graph. We still do not know how well direct capsule retrieval would work if it could read original records through source IDs and keep searching when evidence is missing. ReadAgent appears in the main tables, but the systems differ in other ways, so that comparison does not tell us how much of the gain comes from the graph.

    I suggest a simple comparison on LoCoMo with GPT-4.1-mini: keep the same compressed content, residual addresses, and evidence checks, remove the graph, and use vector and lexical retrieval for both initial and follow-up searches. Comparing macro F1, MultiHop F1, and total query cost under the same source-access, LLM-call, and context budgets would help show whether the graph is worth its overhead.

  2. Experimental costs and reproducibility. It is unclear which calls are covered by the claim of comparable token cost. Appendix A.3.2 separates construction and query costs, but does not clarify whether all input and output tokens for the evidence checks, relation verification, and follow-up search in Section 3.4 are included. Please give costs by stage for one representative configuration, including capsule summarization, and report storage for the graph and recovery mappings.

    The main tables also reuse published baseline results from SimpleMem. I suggest rerunning at least one representative configuration under the same evaluation script. For Figure 3, please clarify whether residual addressing and the ordinary keyword index use the same keyword counts, index budgets, and retrieval settings. I could not find code or a repository link in the supplied materials. An anonymous implementation and experimental configurations would help, including whether semantic links are reused across questions in the main experiments and ablations.

Questions for authors
  1. Could the authors add the comparison in W1 to establish the graph's added value when the same representations and iterative search are retained?
  2. Which calls are included in Token Cost? Could construction and query costs be broken down for one representative configuration?
  3. Do the two lexical indexes in Figure 3 use the same budgets and retrieval settings? Are semantic links reused across questions in the main experiments?
8 comments

We thank the reviewer for the valuable suggestions. We conducted the requested graph-free comparison, broke down costs by stage, reran SimpleMem under our evaluation script, and released the code. Specifically,

  • Graph value (W1/Q1). With the same capsules, residual addresses, evidence checks, and budgets, adding the graph raises macro F1 from 48.05 to 49.72 (MultiHop +2.12), while lowering online cost from 518 to 502 tokens per question and follow-up calls from 150 to 94.
  • Cost accounting (W2/Q2). Token Cost counts all input and output tokens of every online answering call, including follow-up calls. Offline capsule generation and relation verification are reported separately; coarsening reduces offline construction tokens by 54%.
  • SimpleMem rerun (W2). Under our script, SimpleMem reaches 37.99 macro F1 (43.24 published). DuCoMem's advantage holds under both numbers.
  • Index controls and link reuse (Q3). Both lexical indexes share the same 20-term budget and retrieval settings. Verified semantic links are reused within a conversation.
  • Code: The implementation is available in our anonymous repository.

W1a: Relationship to ReadAgent, GraphReader, and STPGC

Role of the graph. I find the motivation for residual addressing reasonable, but I am less clear about what the graph adds. The collapse rules in Section 3.3 draw on STPGC. ReadAgent already uses summaries to locate and read original text, and GraphReader searches a graph and continues when the evidence is insufficient. The paper needs to explain more directly what coarsening and recovery mappings improve over these existing approaches.

Response:

The graph lets DuCoMem reach evidence that is weakly matched to the question but closely related to records already retrieved, and coarsening with recovery mappings makes this graph cheaper to build without losing access to the original records. Compared with the three related methods: i) ReadAgent uses summaries to select pages for rereading, whereas DuCoMem uses the evidence already found and its links to other records to narrow the scope of follow-up rereading. ii) GraphReader builds its graph through LLM extraction over every text chunk, whereas DuCoMem coarsens a low-cost candidate graph and verifies only selected relations, which reduces construction tokens by 53.83% relative to the uncoarsened graph. iii) STPGC's general-purpose coarsening rules do not account for source access, whereas DuCoMem adds source-recovery constraints so that the original records of collapsed nodes remain accessible on demand. The detailed analysis are listed as follows:

i) ReadAgent [B1]. A record with low direct similarity to the question may still contain useful evidence. DuCoMem can locate it through a link to an already retrieved record and use the recovery mapping to read the original text. In the matched graph-free comparison in W1b, both variants use the same capsules, residual terms, and retrieval budgets. Adding the graph reduced the average number of original source records read per question from 60.75 to 59.66 and improved Macro F1 by 1.67 points.

ii) GraphReader [B2]. GraphReader uses an LLM to extract atomic facts and key elements from each text chunk when building its graph. DuCoMem reuses existing memory capsules, builds a candidate graph from their semantic similarity, and coarsens it before using an LLM to verify selected relations. This avoids an additional extraction pass over every chunk and narrows the set of relations to verify. In our LoCoMo construction experiment (W2a), coarsening reduced relation verification token usage by 63.27% and total construction token usage by 53.83% relative to the uncoarsened DuCoMem graph.

iii) STPGC [B3]. STPGC is a general-purpose topology-preserving graph coarsening method. In a memory graph, a capsule that meets the structural conditions for collapse may still point to an original record containing unique evidence. We extend the collapse rules to account for source recoverability and map collapsed capsules to their surviving representative nodes. When retrieval reaches a representative node, the system can use these mappings to read the associated original records on demand.

[B1] Lee et al. A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts. ICML 2024.

[B2] Li et al. GraphReader: Building Graph-based Agent to Enhance Long-Context Abilities of Large Language Models. Findings of EMNLP 2024.

[B3] Wu et al. Scalable Topology-Preserving Graph Coarsening: Concepts and Algorithms. ICML 2026.

W1b / Q1: Added value of the graph over direct iterative retrieval

[W1] Table 5 shows that coarsening helps: macro F1 increases from 48.78 to 49.72, while Token Cost drops from 598 to 502. However, w/o Graph Coarsening still uses the original graph and follow-up search, and w/o Focused Search also retains the graph. We still do not know how well direct capsule retrieval would work if it could read original records through source IDs and keep searching when evidence is missing. ReadAgent appears in the main tables, but the systems differ in other ways, so that comparison does not tell us how much of the gain comes from the graph.

[W1] I suggest a simple comparison on LoCoMo with GPT-4.1-mini: keep the same compressed content, residual addresses, and evidence checks, remove the graph, and use vector and lexical retrieval for both initial and follow-up searches. Comparing macro F1, MultiHop F1, and total query cost under the same source-access, LLM-call, and context budgets would help show whether the graph is worth its overhead.

[Q1] Could the authors add the comparison in W1 to establish the graph's added value when the same representations and iterative search are retained?

Response:

Under the same representations, evidence checks, and budgets, adding the graph improves Macro F1 from 48.05 to 49.72 and MultiHop F1 from 38.16 to 40.28, while lowering total online cost from 518 to 502 tokens per question. The graph improves F1 in all four question categories, with the largest gains on SingleHop (+4.02) and MultiHop (+2.12). It also reduces follow-up answering calls from 150 to 94 and the original source records read per question from 60.75 to 59.66. The detailed results are provided as follows:

1) Configuration of the graph-free and DuCoMem variants. To achieve a fair comparison, we first obtain a DuCoMem variant (keep the same compressed content, residual addresses, and evidence checks, remove the graph, and use vector and lexical retrieval for both initial and follow-up searches). Therefore, we compare both variants on all 1,540 LoCoMo questions with GPT-4.1-mini, where they share the compressed capsules, residual addresses, access to original records through source IDs, and evidence checks. Both variants can search again when the initial evidence is insufficient, under the same limits on source reads, LLM calls, and final context. The graph-free variant combines vector and lexical retrieval for both initial and follow-up search.

2) Source reads, answer quality, and query cost. Table B1 shows that with the graph, Macro F1 rises from 48.05 to 49.72, total online token usage falls from about 518 to 502 tokens per question, and the mean number of original source records read falls from 60.75 to 59.66 per question. Under the same evidence-sufficiency check, the graph-based variant triggers 94 follow-up answering calls, compared with 150 for graph-free variant. Follow-up answering accounts for 15.48 of the 15.99 additional online tokens per question for graph-free variant. Table 5 of the manuscript separately reports the effect of coarsening an existing graph.

The MultiHop gain illustrates the value of finding evidence across records. In one LoCoMo question about Melanie's pets, the graph-based variant retrieves both the earlier record naming Luna and Oliver and the later record about adopting Bailey on its first pass, then answers with all three names. Graph-free variant retrieves only the earlier record and misses Bailey. Across all 282 MultiHop questions, the graph-based variant retrieves every annotated source for 62 questions, compared with 60 for graph-free variant.

Table B1. Graph-guided versus graph-free iterative retrieval on LoCoMo with GPT-4.1-mini.

MetricWith graphGraph-free variantWith graph − graph-free variant
Macro F149.7248.05+1.67
MultiHop F140.2838.16+2.12
Temporal F160.8460.74+0.10
OpenDomain F135.6335.22+0.41
SingleHop F162.1158.09+4.02
All online input and output tokens per question (rounded)502518−16
Original source records read per question59.6660.75−1.09

W2a / Q2: Token accounting and storage

[W2] Experimental costs and reproducibility. It is unclear which calls are covered by the claim of comparable token cost. Appendix A.3.2 separates construction and query costs, but does not clarify whether all input and output tokens for the evidence checks, relation verification, and follow-up search in Section 3.4 are included. Please give costs by stage for one representative configuration, including capsule summarization, and report storage for the graph and recovery mappings.

[Q2] Which calls are included in Token Cost? Could construction and query costs be broken down for one representative configuration?

Response:

The reported Token Cost covers all input and output tokens of every online LLM call, namely the initial answering calls and the answering calls after follow-up retrieval, including the evidence-sufficiency assessment and online relation checks performed within these calls. Offline capsule generation and relation verification are not part of Token Cost and are reported separately. For GPT-4.1-mini on LoCoMo, the online cost is 502 tokens per question, of which answering after follow-up retrieval accounts for 5.2% (Table B2). Offline, coarsening reduces relation-verification tokens by 63.27% and total construction tokens by 53.83% (Table B3), and reduces the storage of the graph, recovery mappings, and source-text inverted index from 5.83 MB to 4.40 MB (Table B4). The detailed results are provided as follows:

1) Online query costs. We use GPT-4.1-mini to answer all 1,540 LoCoMo questions. The input and output counts cover the full online requests and responses, including the evidence-sufficiency assessment and relation checks performed within answering calls. Answering after follow-up retrieval accounts for 5.8% of LLM calls and 5.2% of online tokens.

Table B2. Online LLM calls and token usage on LoCoMo with GPT-4.1-mini.

Online stageLLM callsInput tokensOutput tokensTotal online tokens (per question)
Initial answering (including evidence-sufficiency assessment and online relation checks)1,540712,64519,897475.68
Answering after follow-up retrieval9439,68275826.26
Total1,634752,32720,655502

2) Offline construction costs. We compare construction with and without graph coarsening on one LoCoMo conversation. Both settings use GPT-4.1-mini, the same capsules, and the same relation verification rules. In this comparison, both variants continue relation verification until at least 200 original source records have been covered. As shown in Table B3, coarsening reduces token usage for graph relation verification by 63.27% and total offline construction token usage by 53.83%.

Table B3. Offline memory construction costs on one LoCoMo conversation with GPT-4.1-mini.

Construction componentWithout coarseningWith coarsening
Capsule generation (tokens)35,58235,582
Graph relation verification (tokens)202,86474,508
Total construction (tokens)238,446110,090

3) Storage for the graph and recovery mappings. Across all ten LoCoMo conversations, we use the same lossless JSON format before and after coarsening and verify that the files can be decoded. Table B4 shows that coarsening reduces the size of the graph nodes and edges from 4.74 MB to 3.33 MB. Including the recovery mappings and source-text inverted index, the combined size decreases from 5.83 MB to 4.40 MB.

Table B4. File sizes before and after coarsening across all ten LoCoMo conversations.

Storage componentBefore coarseningAfter coarseningReduction
Graph nodes and edges4.74 MB3.33 MB29.73%
Graph, recovery mappings, and source-text inverted index5.83 MB4.40 MB24.42%
Full runtime JSON files6.96 MB5.53 MB20.45%
Reviewer_MfDFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents an approach for memory compression and retrieval for long context LLM inference. It follows a two step process. First the context is compressed and only the salient tokens are retained in the memory. The residual tokens left out of the memory capsules are stored in additional memory structures. Additionally, a graph structure representing the semantic relation of the context tokens is constructed using vector similarities. The graph is coarsened by deleting nodes and edges that preserve a covering criterion. During retrieval, the residual tokens are used to search over the compressed memory and the key semantic concepts are identified. These concepts are searched over the semantic similarity graph (coarsened version). Once the matching concepts are identified, the graph is locally expanded and the relevant memory capsules are fetched.

Experimental results are presented for the LoCoMo benchmark comparing seven major LLM for the answering task. Ablation is reported for the various components. An improvement of performance is reported with a similar token budget.

Critical strengths

The strength of the paper are:

  1. It improves over state of art memory management techniques for long context inference.
  2. The idea of using residual memory structures for retrieval is interesting and useful.
  3. Maintaining a semantic relation graph might improve relevance of the retrieved memory capsule in most instances.
  4. Initial search over a coarse graph and then locally expanding the search is akin to most database search structures and might provide efficiency while preserving performance.
  5. Experimental results are reported in detail.
Critical weaknesses

In my opinion following are some of the weaknesses:

  1. The memory construction and retrieval process has additional steps and might effect the inference time. Relative benefits of using a more elaborate memory structure vis-a-vis a simpler memory structure in terms of caching effectiveness and inference overhead might not always be in favor of the proposed method.
  2. The assumption that residual tokens stored in additional memory are key for searching over the compressed memory structure might not always be effective. It appears this is just an assumption. I did not find a clear justification for devising this step in the retrieval process.
  3. The effectiveness of the proposed method relies on the quality of the coarsened semantic similarity graph. This graph may not be of good quality, particularly when the past queries are heterogeneous in nature. Similarly, the graph coarsening step may not always lead to a smaller graph that fits in memory.
  4. Although the experimental results are elaborate, some more results illustrating the additional overhead of the proposed system might provide clarity on the effectiveness of the system.
Questions for authors
  1. There are a number of parameters and heuristics used in the memory compression as well as the graph construction phase. For example, what is the semantic similarity threshold used to define an edge in the semantic similarity graph? Some of the details about choosing a value of these parameters are missing in the paper. Authors should do a thorough check on this aspect and add the details to the paper.
  2. Experimental results explicitly studying the additional overhead of the proposed method should be presented and explained.
  3. Error analysis of the proposed method, showing examples of the compressed memory, residual tokens, coarsened graph and the retrieved capsules for some of the LoCoMo data instances may be presented in the results.
  4. Rationale of the proposed two stage memory structure and retrieval, and the complementary nature of the compressed memory, residual tokens and the semantic similarity graph should be discussed in more details.
6 comments

We thank the reviewer for the constructive comments. We added new measurements on LoCoMo with GPT-4.1-mini and will include them in the revised appendix. Our main responses are summarized below.

  • Overhead (W1/W4/Q2). Offline construction costs 114K tokens per conversation, close to LightMem (104K) and about 10× lower than A-Mem. Online cost is 502 tokens per question, 18% below LightMem. The graph adds 0.16 s median retrieval latency, or +0.5% end to end.
  • Residual addressing (W2). This design is supported by experiments, not assumed: removing only residual addressing lowers macro F1 by 1.19 (SingleHop −5.59), and residual cues add entry points for 46% of questions.
  • Graph quality and storage (W3). The graph is built from conversation records, not from past queries, and each new query selects its own entry points. By construction, coarsening never increases storage. On LoCoMo, it reduces storage by 24%.
  • Hyperparameters (Q1), error analysis (Q3), and design rationale (Q4) are detailed below.

W1 / W4 / Q2: Memory construction and query overhead

[W1] The memory construction and retrieval process has additional steps and might effect the inference time. Relative benefits of using a more elaborate memory structure vis-a-vis a simpler memory structure in terms of caching effectiveness and inference overhead might not always be in favor of the proposed method.

[W4] Although the experimental results are elaborate, some more results illustrating the additional overhead of the proposed system might provide clarity on the effectiveness of the system.

[Q2] Experimental results explicitly studying the additional overhead of the proposed method should be presented and explained.

Response:

On LoCoMo, DuCoMem achieves higher Macro F1 than both the lightweight method LightMem [A1] and the higher-cost graph-based method A-Mem [A2] (49.72 vs. 24.63 and 32.58 with GPT-4.1-mini, Table 1), while keeping its overhead close to that of LightMem. We break the overhead into offline construction, online querying, and latency. Offline, DuCoMem uses 114.4K tokens per conversation, close to LightMem (103.9K) and about one tenth of A-Mem (1,195.0K). On a representative conversation, graph coarsening reduces total construction tokens by 53.83%. Online, DuCoMem uses 502 tokens per question, about 18% fewer than LightMem and 80% fewer than A-Mem, and follow-up answering accounts for only 5.2% of these tokens. Its median end-to-end latency is 7.005s, only 0.035s more than our graph-free variant. The detailed experimental results are provided as follows:

Token cost comparison with LightMem and A-Mem. As shown in Table A1, DuCoMem is close to LightMem in offline construction cost and substantially less costly than A-Mem, a graph-based baseline, at both stages. It achieves higher Macro F1 than both baselines while using about 18% fewer online tokens per question than LightMem and 80% fewer than A-Mem. For graph construction, we form candidate edges from capsule embeddings, coarsen the graph, and use the LLM only to verify selected relations. Table A2 reports the construction token costs with and without coarsening. At query time, the graph guides search within local regions, and we search again only when the initial evidence is insufficient.

Table A1. Online and offline token costs on LoCoMo.

MethodOnline token cost (tokens/question)Offline construction cost (K tokens/conversation)
DuCoMem502114.4
LightMem612103.9
A-Mem25201195

Offline construction. We compare the graph without coarsening and the graph with coarsening on one LoCoMo conversation. Both settings use GPT-4.1-mini, the same memory capsules, and the same relation-verification rules. In this comparison, both variants continue relation verification until at least 200 original source records have been covered. Coarsening reduces relation-verification tokens by 63.27% and total construction tokens by 53.83%.

Table A2. Offline memory construction costs for one LoCoMo conversation with GPT-4.1-mini.

Construction metricWithout coarseningWith coarsening
Capsule generation (tokens)35,58235,582
Graph relation verification (tokens)202,86474,508
Total construction (tokens)238,446110,090

Online query costs. We use GPT-4.1-mini to answer all 1,540 LoCoMo questions. Table A3 counts all online LLM input and output tokens, including prompts, retrieved evidence, and generated output. The average is 502 tokens per question. Follow-up answering accounts for 5.8% of LLM calls and 5.2% of online tokens, so most of the online cost comes from initial answering.

Table A3. LLM calls and token costs at query time on LoCoMo.

Online stageLLM callsInput tokensOutput tokensTotal online tokens (per question)
Initial answering1,540712,64519,897475.68
Answering after follow-up retrieval9439,68275826.26
Total1,634752,32720,655502

End-to-end latency. On the same set of LoCoMo questions, the graph-based and graph-free variants have median retrieval latencies of 0.166s and 0.005s, respectively. Their median end-to-end latencies are 7.005s and 6.970s, respectively, a difference of 0.035s.

[A1] Fang et al. LightMem: Lightweight and Efficient Memory-Augmented Generation. ICLR 2026.

[A2] Xu et al. A-Mem: Agentic Memory for LLM Agents. NeurIPS 2025.

W2: Omitted terms as retrieval cues

The assumption that residual tokens stored in additional memory are key for searching over the compressed memory structure might not always be effective. It appears this is just an assumption. I did not find a clear justification for devising this step in the retrieval process.

Response:

Residual addressing is supported by retrieval-level and answer-level evidence rather than assumed: it adds retrieval entry points for 46.04% of LoCoMo questions, and removing only residual addressing lowers Macro F1 by 1.19 points, with the largest drop on SingleHop (62.11 to 56.52). Compression prioritizes the main content of an event, but a later question may use a specific date, place, or name to identify it. For example, when several capsules describe similar trips, the month mentioned in a question can distinguish the intended trip. Residual addressing retains such omitted terms as additional retrieval cues that lead to the corresponding capsule and its original source. The detailed responses are provided as follows:

1) Residual addressing provides additional retrieval entry points on LoCoMo. We compared retrieval with and without residual addressing across all 1,540 LoCoMo questions. For 709 of the 1,540 questions (46.04%), residual addressing added at least one entry point by matching a term in the question to a residual term stored for a capsule. At the answer level, F1 improved for 119 questions (7.73%).

2) Answer quality and online cost with and without residual addressing. Disabling only residual addressing lowers Macro F1 from 49.72 to 48.53. SingleHop shows the largest drop (62.11 to 56.52), consistent with the role of omitted details such as names and dates in locating the single record needed to answer a question. MultiHop and Temporal also decline, from 40.28 to 39.56 and from 60.84 to 60.45, respectively. OpenDomain rises from 35.63 to 37.60. Disabling residual addressing also increases follow-up answering calls from 94 to 103. These added calls help explain why total online LLM input and output tokens rise from 502 to 505 per question.

3) An omitted month can help locate the relevant record. One LoCoMo question asks where Joanna went in July 2022. The compressed capsule omits “July,” which the question uses to identify the trip. Retaining “July” as a residual cue helps the system retrieve the original record and answer “Woodhaven.” With residual addressing disabled, the system misses that record.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We sincerely thank all three reviewers for their constructive feedback, which helped us improve the manuscript. We were encouraged that the reviewers found residual addressing useful for retrieval (MfDF, jfxp, SJXP) and recognized the value of distinguishing content preservation from retrievability (jfxp, SJXP). Reviewer SJXP also highlighted the gains over SimpleMem across all seven answering models on LoCoMo, the results on LongMemEval-S, and the supporting ablations and compression analysis.

We have revised the manuscript in response to the reviewers’ suggestions. The main changes are summarized below.

1. Clarifications on the method and implementation

  • Two-stage design and component roles. (MfDF) We explain the division between offline construction and online retrieval, and how compressed capsules, residual terms, and the graph provide complementary access to evidence (see Appendix B.1.2 in the revised version).
  • Retrieval settings and link reuse. (MfDF, jfxp) We specify the matched term budgets and retrieval settings for the two lexical indexes in Figure 3(a), clarify the construction and search parameters, and describe semantic-link reuse across questions within a conversation (see Appendices A.3.1, B.1.3, and B.1.4 in the revised version).
  • Relationship to existing methods. (jfxp) We explain how DuCoMem’s graph-guided source recovery, construction before relation verification, and source-recovery constraints relate to ReadAgent, GraphReader, and STPGC (see Appendix B.2 in the revised version).

2. Additional experiments and measurements

  • Residual addressing and graph-guided retrieval. (MfDF, jfxp, SJXP) We separately removed residual addressing and compared graph-guided retrieval with graph-free iterative retrieval under matched budgets. Both variants reduce Macro F1 and increase online token cost (see Appendix A.4.1, Table 7 in the revised version).
  • Comparison with SYNAPSE. (SJXP) We evaluated SYNAPSE using the same answering model and scoring script. DuCoMem achieves higher F1 in all four question categories while using fewer online tokens per question (see Appendix A.4.2, Table 8 in the revised version).
  • Construction and query costs. (MfDF, jfxp) We report offline and online token costs, LLM calls, retrieval and end-to-end latency, and storage. The coarsening comparison uses the same source-coverage stopping criterion for both variants, allowing us to compare verification costs at a shared coverage target (see Appendix A.4.3, Tables 9–12 in the revised version).
  • Source recovery and answer analysis. (MfDF) We trace a query through a collapsed capsule and its representative node to the original evidence, and examine the difference between the generated date and the month-level reference answer (see Appendix A.4.4, Table 13 in the revised version).

Thank you again for your time and feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.