Agent Memory: What to Store, How to Compress, and Prevent Staleness

Characterize effective long-term memory design for LLM-based AI agents by specifying what categories of state to store (episodic, semantic, procedural), deriving compression and summarization methods that preserve critical constraints, and establishing safeguards that prevent stale or low-quality memory from dominating subsequent decisions.

Background

Long-horizon tasks require memory beyond the context window, but naïve designs risk contradictions, prompt injection persistence, or decision bias from outdated information. Deciding which memory types to maintain, how to compress them efficiently, and how to enforce provenance and quality is essential for reliability.

These design choices interact with budgets (tokens, retrieval fan-out) and safety. Establishing principled policies for memory writes and updates, along with verification and provenance tracking, is necessary to maintain consistency and robustness in real deployments.

References

Retrieval-augmented generation is a strong baseline, but open questions include what to store (episodic vs. semantic vs. procedural memory), how to compress and summarize without losing critical constraints, and how to prevent stale or low-quality memory from dominating decision making.

— AI Agent Systems: Architectures, Applications, and Evaluation  (2601.01743 - Xu, 5 Jan 2026) in Section 7.2 (Long-Term Memory, Context Management, and Continual Improvement)

The study does not determine whether the model ignored the memory, weakly weighted it, or corrected an intermediate bias.

— Procedural Memory Under Change: Reuse and Interference in Controlled Web Tasks  (2609.09774 - Cao, 9 Sep 2026) in Introduction, final paragraph

The remaining question is which additional conditions turn applicability mismatch into observable, memory-caused error.

— Procedural Memory Under Change: Reuse and Interference in Controlled Web Tasks  (2609.09774 - Cao, 9 Sep 2026) in Abstract; Discussion, final paragraph; Conclusion

These approaches raise a common unresolved question: which information should be remembered, recomputed, or invalidated after an action?

— AI for Games in the Foundation Model Era  (2609.16679 - Luo et al., 15 Sep 2026) in Section 9.2, “Model Players and Games: Persistent World State and Reliable Player Models”

Additionally, managing graph decay---pruning outdated nodes, resolving contradictory constraints, and handling intent pivots---remains an open challenge for long-term deployment.

— CogChat: Knowledge Graph-Augmented Conversational AI with Heterogeneous Graph Transformer for Cognitive Grounding in Design Generation  (2608.13216 - Choi et al., 13 Aug 2026) in Section Discussion ("Limitation")

Length breaks the graph. A graph that adds a node per object instance and an edge per relation per frame is intractable over the minutes-to-hours horizons of real egocentric activity. AMEGO's semantic-free tracklets and FocusGraph's compact textual graphs are the only serious answers so far, and neither offers a general account of how to compress, forget, and re-instantiate graph structure over long horizons.

— Vision-Language Models for Egocentric Video: From Hand-Object Interaction to Embodied AI  (2608.18671 - Zamani et al., 19 Aug 2026) in Section 9.6, “Open Challenges in Graph-based Reasoning” (Sec. graph-challenges)

This raises several questions: how much historical context is needed for reliable interaction, what should be stored or compressed into memory, and how can robots maintain beliefs over partially observed environments without sacrificing control responsiveness? Developing methods that enable robots to incorporate long-term history into decision-making remains an open challenge, and addressing it will be critical for enabling robots to operate reliably in dynamic and continuously evolving environments.

— Toward Unified Robot Learning: Bridging Representation, Vision-Language-Action, and World Models  (2609.03927 - Mehta et al., 3 Sep 2026) in Section 3.4, Long Context Understanding

When should episodic memory be trusted? Past experience may become invalid under changes in sensing, policy, traffic rules, or operating conditions, whereas excessive forgetting may cause repeated failures. Memory provenance, confidence decay, conflict tracking, versioning, and revalidation are therefore central challenges for long-term safety memory.

— Rethinking World Models for Safety-Critical Embodied Systems  (2609.03774 - Ma et al., 3 Sep 2026) in Section “Open challenges and outlook,” subsection “Open challenges”

Every distilled prior is critic-derived rather than ground truth, and retrieval quality, staleness, and the propagation of a wrong prior remain open.

— VeriPhy: Agentic Physical Reasoning for World Model Evaluation and Refinement  (2609.03153 - Xu et al., 2 Sep 2026) in Section 6, Limitations

The store offers contest, demote and supersede each round and no compared policy uses them, so correction is untested rather than absent.

— A Benchmark and Diagnostic Study of Epistemic Admission in Shared Agent Memory  (2609.30813 - Li et al., 25 Sep 2026) in Appendix, Section “Limitations” (Section~\ref{app:limits})

These methods improve memory organization, revision, and protection but leave open when statements should enter memory.

— AgentMemGate: Addressing Speculation Contamination in Conversational Assistant Memory  (2610.07707 - Sharma et al., 6 Oct 2026) in Related Work, subsection “Write-time control”

This is a precision-recall tradeoff inherent to independent fact evaluation, not an artifact of our specific threshold, and motivates chain-aware admission that models inter-fact dependencies at formation time as the primary direction for future work.

— Gated Memory: Admission-Controlled Memory Formation for Conversational AI  (2610.11270 - Saraswat et al., 8 Oct 2026) in Section 5, paragraph “Multi-Hop regression”

What remains open is how an agent should decide when new evidence replaces an older state, when both should coexist, and when uncertainty should be preserved rather than resolved.

— Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory  (2610.19366 - Guo et al., 14 Oct 2026) in Section 5.2.1, “Maintaining Up-to-Date Memory” (Section sec:open-up-to-date-memory)

Our evaluation is limited to question answering over static corpora, and maintaining anticipatory memories as documents change remains an open problem.

— Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries  (2610.22273 - Singh et al., 15 Oct 2026) in Section 4, Discussion, Limitations, and Future Work, p. 11

The unresolved issue is distinguishing valid current information from historical records.

— Physical AI: A Survey of Models, Learning, Systems, and Open Challenges  (2610.21171 - Gao et al., 15 Oct 2026) in Section 10.4.1, Maintaining Task-State Estimates and Memory

These operations motivate a shared research question: under a limited context budget, what information should be carried forward to support subsequent decisions?

— Mosaic: Multi-Session Agentic Reinforcement Learning  (2610.22174 - Xu et al., 15 Oct 2026) in Discussion: Future Directions for Multi-Session Agentic RL, paragraph “Cross-session state transfer as a central learning objective”

We do not evaluate whether learned assembly helps with other memory types, such as semantic memory (facts and rules) or episodic memory of a particular user's history, where the interaction between selected items and their order may differ.

— MemComposer: Learning to Assemble Memories for LLM Agents via Reinforcement Learning  (2610.21091 - Xie et al., 15 Oct 2026) in Section Limitations, paragraph “Memory type”