acceptodds
Withdrawn from ICLR 2027

When External Memory Becomes Agent State: Recorded Lineage and Controlled Reuse in LLM Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Persistent-memory agents can rewrite external records into their own long-term state. A terminal success rate does not reveal whether an external record was merely shown, used in a decision, copied into a new record, inherited by later writes, or used again. We study recorded lineage: a committed record is linked to an external seed only through parent identifiers declared by the model and validated by the evaluator. We analyze this operational relation separately from content correctness, retrieval, and outcomes. Across three deterministic studies, ordinary writeback created 48 seed-descended records; disabling all writeback created none but reduced task success from 5/10 to 0/10. Two selective policies then blocked all recorded descendants. Lineage-selective admission nevertheless committed 36 DeepSeek and 3 Kimi writes without recorded seed ancestry, whereas exposure-triggered admission committed none. On the original task horizon, the prespecified operational comparison was near floor and did not favor lineage selection. Finally, when a retained notice-derived record was scheduled for later use, lineage selection achieved 67/120 versus 1/120 successes on DeepSeek and 72/120 versus 6/120 on Kimi. Scrambling only the record’s task-relevant content reduced success to 4/120 and 9/120. Together, the studies separate lineage formation, write-policy selectivity, and consequence after scheduled reuse. Code and anonymized data are available in an anonymous artifact.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, persistent memory, memory provenance, recorded lineage, write-boundary control, agent safety, long-term memory

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

The paper studies provenance tracking for persistent-memory LLM agents using model-declared parent identifiers. It compares global write blocking, exposure-based blocking, and lineage-based blocking across synthetic tasks. While lineage-based filtering preserves more writes, the main operational evaluation shows little meaningful improvement. A separate controlled-reuse experiment reports large gains when relevant memory is explicitly retained and retrieved.

Critical strengths

The paper makes a useful distinction between memory exposure, recorded ancestry, and downstream behavior. The attempt to evaluate these stages separately is understandable, and the experiments include controlled comparisons and a content ablation.

Critical weaknesses

In general, the paper is extremely difficult to read. I feel this is a more like an ai-generated tech report, than a research paper, there is no clear research gap, research contribution, and the paper presentation needs significant improvements.

The main mechanism is essentially provenance tracking through explicitly declared parent IDs. The proposed policy ordering follows directly from the definitions, and it is unclear what substantial technical contribution is made beyond existing provenance and information-flow tracking approaches.

Both studies also lack descriptions and hard to follow. Study 2, the more natural operational evaluation, shows little improvement in task success or invalid-event rates. The large gains in Study 3 come from a constructed setting where one policy retains a task-relevant record, the other removes it, and the agent is subsequently required to use that information. This demonstrates the benefit of retaining useful information, but provides limited evidence that the proposed approach improves realistic persistent-memory agents.

Questions for authors
  1. What is the technical novelty over existing provenance tracking and lineage-based enforcement methods such as MemLineage?
  2. Why should Study 3 be considered evidence of practical utility when the downstream task is explicitly constructed to depend on the record preserved by one policy?
  3. How would the approach handle incomplete or incorrect parent declarations, especially when an agent incorporates information without reporting its source?
Reviewer_o6GeOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how external records become persistent memory for LLM agents through write-back. It tracks record lineage using parent-record links declared by the model and checked by an evaluator. It also compares lineage-based and exposure-based policies for deciding which records can be written to memory.

Critical strengths

The writing shows strong signs of AI use, the experiments are unconvincing, and the paper has no clear strengths.

Critical weaknesses
  1. The lineage mechanism relies on the model’s own reports, but their reliability has not been tested.
  2. The key difference between the policies in Study 3 is explicitly built into the experimental instructions. Study 3 is closer to a controlled demonstration of the mechanism.
  3. The largest gains come from specially constructed scenarios.
Questions for authors

See weaknesses.

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

This paper shows how an external seed record (m0m_0) read by an LLM agent can be written back into its own persistent state, and proposes a lineage-based write-boundary method (admit_lineage) to track and control this propagation across five stages (exposure, declared use, writeback, recorded ancestry, and downstream actions). To demonstrate the effectiveness of the proposed method, the authors perform experiments across diverse scenarios using DeepSeek and Kimi: they examine the baseline impact of globally disabling writeback (Study 1), compare selective containment against exposure-triggered blocking (admit_exposure) on the original task horizon (Study 2), and evaluate a scheduled reuse scenario where the agent must later rely on a retained parent-free record, showing a +55.0 percentage point advantage for admit_lineage (Study 3).

Critical strengths

(S1) Five-Stage Decomposition and Deterministic Evaluation: Rather than just looking at terminal task success, the authors break down the transition from retrieval exposure (RtR_t) to recorded ancestry (Gt=(Mt,Et)G_t = (M_t, E_t)) and show in Study 1 that having seed ancestry (48 committed descendants) is not the same as being semantically false (12 deterministically false, while 36 remain consistent with targeted claims). In addition, the authors build a deterministic evaluator without relying on an LLM-as-a-judge and openly report the null operational result on the original task horizon in Study 2 (p=1.0p=1.0), which makes the evaluation setup clear and reproducible.

Critical weaknesses

(W1) Limited practical utility of blocking Writes under continued exposure: Both admit_exposure and admit_lineage assume that the evaluator already tracks the seed family (St\mathcal{S}_t) rooted at m0m_0. However, keeping a stale m0m_0 visible in retrieval (m0∈Rtm_0 \in R_t) while only blocking candidate writes at the write boundary does not help much in practice. In Study 2 (Table 13), simply excluding m0m_0 at retrieval time (No retrieval, m0∉Rtm_0 \notin R_t) led to 100% terminal task success (DeepSeek 25/25, Kimi 15/15). In contrast, Lineage-selective achieved only 24.0% (6/25) and 33.3% (5/15) success, which is similar to the Seed uncontained baseline (5/25 and 7/15), because the agent is still misled by the exposed m0m_0. The authors do not provide a convincing practical reason for why a system that already tracks m0m_0 would only control the write boundary instead of excluding m0m_0 from retrieval once it becomes stale.

(W2) Difficulty of applying the method to existing memory stores: While the paper is motivated by practical agent memory management, admit_lineage is hard to apply to real-world databases that are already populated. For ancestor closure (Anct(w)\text{Anc}_t(w)) to work properly, every committed record in MtM_t must record explicit parent_memory_ids (Pt(w)P_t(w)) starting from t=0t=0. In existing RAG or memory systems where past records do not have recorded lineage edges (EtE_t), if even one previously derived record (m1m_1) lacks a recorded path from m0m_0, any future write citing m1m_1 will satisfy Anct(w)∩St=∅\text{Anc}_t(w) \cap \mathcal{S}_t = \emptyset and completely bypass the containment policy.

(W3) Reliance on self-declared parent links and prompt Instructions in Study 3: Since the evaluator only checks whether declared parents are currently visible (Pt(w)⊆RtP_t(w) \subseteq R_t), it cannot tell if a model secretly copies stale values from a visible m0m_0 while submitting an empty parent list (Pt(w)=∅P_t(w) = \emptyset). More importantly, the +55.0 percentage point gap in Study 3 (Table 14) comes from the Stage s3 prompt (Appendix H.2), where the authors explicitly tell the model: Do not carry values from the visible memory record into this write, and set parent_memory_ids to an empty list. Because of this direct instruction, Study 3 does not show whether models can naturally separate and declare parent sources (wpubw_{\text{pub}}) on their own without prompt guidance.

Questions for authors

Please see the critical weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.