acceptodds
Under review as a conference paper at ICLR 2027

Agentsensus: Consensus-Compressed Shared Memory for Multi-Agent Story Worlds

Abstract

A agentic story world is a dynamic system simulating who learned what, when, and from whom — yet the standard design gives each character a private memory stream. A shared event is therefore stored once per witness, large duplication will be incurred in terms of storage. We present Agentsensus, a story-world simulation framework in which there is an unified long-term memory. Records of the same event merge into one owned by all its witnesses, and semantically relevant memory records are linked. We evaluate on four worlds — two classical Chinese novels, Hamlet, and a real-world conflict timeline — run for 40 to 80 rounds against three per-character memory designs under an equal-granularity protocol. Agentsensus writes 22–44% fewer entries than the closest baseline and is the only design whose memory becomes shared (14–28% of records held by more than one character, some by 10) and linked (94–99%), at judged simulation quality indistinguishable or even better than the baselines. An ablation attributes this to the merge itself: disabling it multiplies the store by 3.1× and takes sharing to exactly zero. Sharing also compounds with the horizon rather than saturating early, rising 6%→9%→14% as one world is re-run at 10, 20 and 40 rounds.

Primary area:
foundation or frontier models, including LLMs
Keywords:
multi-agent systems, memory, large language model agents, agent simulation, shared memory, narrative generation

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

Agentsensus merges similar event memories into shared records and expands retrieval through record links to reduce storage while supporting multi-agent story continuation.

Critical strengths

Agentsensus writes much fewer simulation entries than the baseline.

Critical weaknesses
  1. Keeping shorter text removes useful facts. Appendix B finds lost names in 44 of 147 recoverable merges, including 18 losses of the ordering party. This selected subset does not estimate the overall rate. The quality scores does not establish preserved authority information; compare facts before and after merging.

  2. It reports five merges that combine distinct events, with no demonstrated way to undo them. Shared records can then spread the wrong account through later retrieval. One wrong merge affect later decisions. Preserve original records and test detection, reversal, and recovery from false merges.

  3. Consensus retrieves a median of 32 records versus five for a baseline. Appendix B’s uncapped expansion could replace details lost through compression. Similar output quality therefore does not establish preserved information. The comparison need equal retrieved-token budgets and checks of whether extra records compensate for missing facts.

  4. It compares backends generating different continuations and later memory deposits. A shared atomizer does not guarantee identical incoming events. Fewer stored records could reflect both compression and different activity. The live results remain useful, but the saving need a direct test on shared input. Replay identical deposits through each backend and compare retained facts, final records, and storage bytes.

  5. It rewrites logs into screenplays for three quality measures; goal pursuit uses raw logs. Faithful rendering is requested but not independently verified. The rewrite could hide memory errors. The screenplay scores do not validate both stages.

  6. A agentic story world to An agentic story world. an unified long-term memory to a unified long-term memory.

Questions for authors

Please refer to the weaknesses.

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

The paper listed three weaknesses in agentic story world study: witness-multiplied redundancy, fragmented, divergence-prone world knowledge, and memory management that agents never perform. To solve these research questions, the authors propose Agentsensus, a framework that keeps all characters’ memories in one store. When different characters record the same event, their records are merged into one record. In addition, short statements that come from the same original memory are linked to each other. They compare this framework with three other baselines and experiment results show that Agentsensus can write down fewer entries. On other aspects, such as continuation quality, there is no clear difference between Agentsensus and the baselines.

Critical strengths

The authors provide detailed code for review and show how the whole research project was built. This makes the work easy to reproduce and the experiment process more transparent. Although there are some Chinese docs mixed in it. For example, in the README.md file. I would appreciate it if all of them are written in English.

The authors were very honest in reporting their experimental results, making no attempt to discard unfavorable findings. The definition of three structural problems is clear.

Critical weaknesses

1.Evaluation and significance

Using well-known novels for evaluation carries a risk of data contamination. The models have probably seen the full text of these works during training, so the grounding and trajectory scores, which compare the simulation with the original works, may mainly reflect what the models already know rather than the effect of the memory design. This concern is hard to assess because the paper does not state which models are used. The AI Use Statement says that the models and versions are given in section 3, 4 and Appendix A, but I could not find them there. The embedding model and the judge model are not specified either.

The main empirical result is that Agentsensus writes fewer memory entries. The paper also reports sharing and linking ratios, but these are zero for every baseline by construction (Table 2), so they cannot show an advantage over the baselines. What is missing is evidence that fewer entries bring a real benefit. The paper’s own results point the other way: continuation quality is about the same as the baselines, write latency roughly doubles because of the extra LLM judge call, and the largest store has only 7,025 rows, where storage is not a bottleneck.

I would suggest choosing metrics that match the goal of the paper. If the work is about agent memory, QA-based tests could check whether each character can recall the right information, and the accuracy should be reported. If the work is about agentic story generation, metrics for literary quality should be used.

  1. Novelty and Contribution

The contribution of this paper is stated as making ownership a property of the record, not the store by the authors. However, most of the design and component concepts are discussed in other papers. The authors also admit that they mostly use standard components (atomization, an embedding pre-filter and an LLM equivalence judge) in their method. Therefore I think the authors should introduce more novel designs to make the contribution a conference-level paper.

  1. Some other aspects

The paper writing should be improved, as the grammar mistakes (e.x. A agentic story) can be found even in the abstract. It is also recommended to cite more recent works. All papers in related work part come from years before 2026. Papers like "EvoSpark: Endogenous Interactive Agent Societies for Unified Long-Horizon Narrative Evolution" (ACL 2026) are also related to this work. These aspects did not influence my review score, but improving them can make the paper better.

Questions for authors

1.Figure 2b describes G-Memory as a two-tier graph kept per owner. The original G-Memory paper uses a three-tier hierarchy (insight, query and interaction graphs). I would like to know how this baseline was adapted, and whether it is really a faithful reimplementation.

2.The merge rule treats two records as equivalent. After sharing the consensus over different characters, how can we protect different characters' diverse opinions towards the same thing? In a story, different characters may see the same event differently, or one may hold a false belief. Has it been tested in settings with information asymmetry, such as deception or secrets? This is important for the literary quality of a work.

3.Please disclose the LLM models used in the experiment, or if I have overlooked this, could you point out where this information is mentioned in the text?

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

This paper proposes Agentsensus, a shared memory framework for multi-agent story worlds. Instead of storing separate copies of the same event for each character, it merges similar memories into a single record while keeping track of which characters know the event. It uses an LLM to decide whether memories describe the same event and links related memories to support retrieval.

The method is evaluated on four story worlds, including Three Kingdoms, Dream of the Red Chamber, Hamlet, and a Russia–Ukraine timeline. Compared with three baseline memory systems, Agentsensus stores 22–44% fewer new memory records while achieving similar story quality. The paper also studies memory sharing, incorrect merges, information loss, and retrieval costs.

Critical strengths

Sharing memory records while keeping track of which agents know each event is a useful idea. The experiments use a common memory extraction process across methods and include an ablation without merging.

The paper provides useful analysis beyond memory size, including story quality, incorrect merges, lost information, and retrieval costs. It also discusses several limitations of the approach.

Critical weaknesses

It is unclear whether merging memories preserves all important information. Among 147 examined merges, 44 lost character names and five incorrectly combined different events. The experiments also use only one simulation run per setting, and Agentsensus retrieves much more context than the baselines.

The released code appears to differ from the experimental description. Generative Agents seems to use shared evidence during reflection, while G-Memory may not fully restore its graph state from checkpoints. These differences could affect the fairness of the comparisons.

Questions for authors

Can the authors measure how often merging loses important event details or ownership information? Can they also repeat some experiments with the same retrieval context budget across methods?

Can the authors clarify which code version was used for the experiments and verify that the baseline implementations correctly preserve their memory state and information boundaries?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.