acceptodds
Under review as a conference paper at ICLR 2027

MemMooring: What Memory Keeps and Recalls Over Time

Abstract

A sentence about a life already carries the shape of the world it is about: the people and places, what changed, in what order, and what has not yet happened. Language is a structured abstraction of the world and of its changes, so a memory built from conversation can recover what was talked about rather than the words used. We present MemMooring, which rebuilds a conversation as a world representation that grows over time and is corrected in place across sessions rather than overwritten. Events retain world time, recording time and their source utterances; the states they change are kept as versions carrying intervals of validity and of recording. The evaluated stores are extracted with Ollama qwen3.6:latest (36.0B MoE, GGUF Q4_K_M, 262,144-token context); answering uses MiniMax-Text-01 at temperature 0, with DeepSeek-chat and GPT-4o as separate judges. The representation is complete enough to be read in two directions: a horizontal cut returns everything about one person at one moment — the facts then in force, the events under way or still consequential, and the plans already known whose time still lay ahead — and a longitudinal cut follows one matter from the decision that opened it through its recorded steps to where it stands. We evaluate the implemented storage, retrieval and answering paths on ten LoCoMo conversations. Under that protocol MemMooring reaches 86.6% and 83.1% against 85.5% and 82.2% for current Mem0. On the 181 questions that name a calendar date, adding a channel that orders events by the distance between their stored interval and the named day raises accuracy from 77.3% to 83.4% and from 75.7% to 81.8% at an unchanged limit of thirty records, and raises strict evidence coverage from 78.3% to 88.9%; on the 1356 questions that name no date the two retrievals return the same 81.4% coverage. The two cuts are shown as transcript-linked worked examples over one conversation; their set and process accuracy are not annotated and are not claimed.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
long-term memory, conversational agents, temporal memory, episodic memory, state tracking, retrieval, memory systems, LoCoMo

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_LQbWOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents MemMooring, a memory system for conversational AI agents that preserves historical events and tracks how user states evolve over time. The authors evaluate MemMooring on LoCoMo and report 86.6% accuracy vs. 85.5% for Mem0, the difference is small and is not statistically significant.

Critical strengths

none

Critical weaknesses

W1. The text font doesn’t follow the ICLR format.

W2. The presentation is very weak: The paper is hard to read, and the text in the figure is too small, which is impossible to see when printed.

W3. The paper’s main conceptual contributions are the horizontal cut and longitudinal read, which aim to reconstruct historical world states and trace the evolution of conversation. but none of them is evaluated. The authors provide only two qualitative case studies from a single conversation (Section 4.5) and quantitative experiment only focus on general response quality

W4. This paper possess limited methodological novelty over existing memory management for long-term conversations. For instance, the concept of event → state is formulated in [1] . Although the authors position horizontal and longitudinal reads as novel, the paper does not clearly establish which capabilities cannot be achieved using existing temporal memory architectures.

W5. The number of baseline methods in the experiment is too small (2; No memory and Mem0), which doesn’t quite fit the common practice in ICLR papers. The authors did not even include the naive summarization-retrival baseline.

[1] Kai Tzu-iunn Ong, Namyoung Kim, Minju Gwak, Hyungjoo Chae, Taeyoon Kwon, Yohan Jo, Seung-won Hwang, Dongha Lee, and Jinyoung Yeo. 2025. Towards Lifelong Dialogue Agents via Timeline-based Memory Management. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 8631–8661, Albuquerque, New Mexico. Association for Computational Linguistics.

Questions for authors

Please refer to W3-W5 thank you.

Reviewer_6U47OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes MemMooring, a conversational memory that stores dialogue as entities, events, and versioned states. Each state carries a valid interval and a recorded interval. Spoken time phrases are kept verbatim with their utterance day and a derived interval. Writes append versions instead of overwriting, and conflicting accounts are kept with their sources. Two reads are defined: a horizontal cut H(t), which returns a person's situation at a given moment, and a longitudinal read L(M), which traces one matter over time.

On 1540 LoCoMo questions, MemMooring scores 86.6% and 83.1% under two judges, against 85.5% and 82.2% for Mem0. On 181 date-naming questions, an added interval channel raises accuracy from 77.3% to 83.4% and from 75.7% to 81.8%. H(t) and L(M) are shown only as worked examples.

Critical strengths

Keeping the spoken phrase, the utterance day, and the derived date separately makes every date recomputable. Conflicting accounts are preserved with their provenance. All systems share the answering model, the prompts, the judges, and a strict one-day date tolerance. The date subset is rule-defined, and paired tests are reported. The authors openly state that the Mem0 comparison is not significant.

Critical weaknesses

All systems share the answering model, the prompts, the judges, and a strict one-day date tolerance. The date subset is rule-defined, and paired tests are reported. The authors openly state that the Mem0 comparison is not significant. The gain over Mem0 is not significant (p = 0.27 and 0.43). The only significant result is a within-system ablation on 181 questions. Mem0 is never given the same date-distance channel, although a timestamp works as a point interval. So the gain is not shown to come from the bitemporal structure. The evaluated stores let the LLM resolve dates, which contradicts the paper's own write contract. Extraction and slot accuracy are not measured, although slot errors are called the dominant failure. No temporal memory system, such as Zep or Engram, is compared. The latency baseline, a linear JSON scan, is weak. The manuscript appears unfinished:

  • Only about 9 of 50+ references are cited in the text.
  • The reference list is unsorted and contains internal notes such as "[verified: …]".
  • Section 4.2 repeats a sentence.
  • Cross-references and the Figure 1 caption are wrong.
  • The prompts are not included.
Questions for authors
  1. If Mem0's records are ranked with the same date-distance channel, treating timestamps as point intervals, does the gain on the 181 date-naming questions remain?
  2. Can you provide a quantitative evaluation of H(t) and L(M)? For example, set precision and recall against annotated cuts, trajectory accuracy, or results on StateMemBench or LongMemEval.
  3. If the stores are rebuilt with the deterministic time resolver, as the write contract specifies, how do the results change? What are the extraction and slot-assignment accuracies?
  4. How does MemMooring compare with an existing temporal memory, such as Zep or Engram, under the same protocol?
Reviewer_e4BGOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents MemMooring, a conversational memory built from entities, events, versioned states, and provenance links, with separate valid-time and recording-time intervals. It supports a horizontal reconstruction of the world at a chosen moment and a longitudinal trace of an undertaking. MemMooring is statistically comparable to Mem0 in overall QA accuracy on LoCoMo benchmark.

Critical strengths
  1. The bitemporal design is conceptually useful.

  2. The evaluation carefully separates storage, retrieval, and answering.

  3. The interval channel gives a consistent empirical improvement on the date-naming subset under both judges, increases strict evidence coverage, and leaves the no-date control unchanged. The paper also reports the index construction costs rather than only lookup latency.

Critical weaknesses
  1. The overall QA is not significantly better than Mem0, and the reported gains concern the RAG retrieval path. The current evidence therefore does not establish that the proposed event–state store improves the two structured reads that distinguish it.

  2. The central representational contribution is not quantitatively validated. The horizontal and longitudinal reads are demonstrated on only one worked conversation, without measurements of completeness, correctness, provenance fidelity, or as-of consistency.

  3. The retrieval ablation does not isolate the benefit of the proposed representation. Embedding-only versus embedding-plus-interval retrieval shows that date-aware ranking helps; it does not show that dual time axes, versioned states, or the event–state schema are necessary.

  4. The paper does not show that dual time axes, versioned states, or the event–state schema are necessary. In particular, the claim that “a timestamped record admits no such ordering” (line 308) seems incorrect: the absolute distance between a timestamp and a query date is well-defined. Perhaps some additional ablations are needed here.

Questions for authors

Most of my concerns and questions are written in the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.