Skip to content

Reading List

Dated snapshot — verified 2026-08-27

Primary sources age better than tool pages, but the frontier moves. Where a survey is superseded, prefer the newer one.

Papers, not products. Read the surveys first — they are the cheapest way to get a map — then follow their citations into whichever stage you are working on.

Start here: surveys

  • Memory in the Age of AI Agents (arXiv 2512.13564) — the broad map, and the source of the extract → store → retrieve → evolve framing this course uses. Accompanied by a maintained paper list at Shichun-Liu/Agent-Memory-Paper-List.
  • Rethinking Memory Mechanisms of Foundation Agents in the Second Half (arXiv 2602.06052) — a more opinionated survey; good on open problems.
  • Always-On Agents: Persistent Memory, State, and Governance in LLM Agents (arXiv 2606.30306) — the governance side, including why deletion has to cascade.
  • AI Meets Brain: Memory Systems from Cognitive Neuroscience to Autonomous Agents (arXiv 2512.23343) — read it for where the biological analogy genuinely constrains design, and where it is decoration.

The write path

  • Belief Memory: Agent Memory Under Partial Observability (arXiv 2605.05583) — treating stored facts as beliefs with confidence rather than assertions.
  • When Does Belief-Based Agent Memory Help? Reliability-Conditional Updating and Provenance-Capped Poisoning Defense (arXiv 2606.22030) — the paper to read before designing an update rule; also the clearest treatment of why provenance is a security control.
  • Less Context, More Accuracy: A Bi-Temporal Memory Engine for LLM Agents (arXiv 2606.09900) — bi-temporal modelling with an empirical case that a lean retrieved context beats full history.

Evolution, consolidation, forgetting

  • Control-Plane Placement Shapes Forgetting: An Architectural Study of Agent Memory Across Thirteen System Configurations (arXiv 2606.15903) — where the forgetting decision lives changes what gets forgotten.
  • Governing Evolving Memory in LLM Agents: Risks, Mechanisms, and the Stability and Safety Governed Memory (SSGM) Framework (arXiv 2603.11768) — drift, runaway self-reinforcement, and decay functions.
  • How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior (arXiv 2505.16067) — what actually happens as stored experience accumulates.

Deletion and privacy

  • Agentic Unlearning: When LLM Agent Meets Machine Unlearning (arXiv 2602.17692) — introduces information backflow: parametric residue regenerating a "forgotten" fact, which is then written back to memory, reversing the deletion. Read before you believe a deletion is complete.
  • MEMOREPAIR: Barrier-First Cascade Repair in Agentic Memory (arXiv 2605.07242) — repairing derived artifacts after a deletion.
  • Forgetful but Faithful: A Cognitive Memory Architecture and Benchmark for Privacy-Aware Generative Agents (arXiv 2512.12856).

Evaluation

  • LongMemEval-V2: Evaluating Long-Term Agent Memory Toward Experienced Colleagues (arXiv 2605.12493).
  • StreamMemBench (arXiv 2606.14571) — streaming evaluation, closer to how a deployed system is actually exercised.
  • ImplicitMemBench (arXiv 2604.08064) — measures unconscious behavioural adaptation, i.e. whether memory changed behaviour rather than recall.

How to read this literature

Three habits worth adopting. Check whether a paper evaluates the write path or only recall — most measure recall and claim memory. Check whether reported gains survive a change of backing model. And check the contradiction handling specifically: a system can look excellent on every benchmark here and still never retire a stale fact, because most benchmarks do not ask.