Skip to content
Research Insights Made Simple
Research reviewResearch Insights Made Simple

Dynamic agent memory

Retaining, revising and testing experience through a CI example

/ Research Insights Made Simple · Agent harnesses 16/19 · Dynamic memory

Slide contents

  1. 1. Dynamic agent memory

    Retaining, revising and testing experience through a CI example

  2. 2. Memory needs four different checks

    Record experience, select current evidence, apply it, forget on request

  3. 3. Continuing work requires selecting experience

    Handoffs, agent anatomy, evals and skills supplied parts of the answer

  4. 4. Procedure and experience answer different questions

    One agent reads a procedure, current test output and past observations

  5. 5. A successful fix can become stale

    One project: failure, verification, upgrade, another task

  6. 6. A-MEM organizes linked notes

    Wujiang Xu and colleagues · Rutgers · NeurIPS 2025

  7. 7. A note stores content and interpretation

    Original content gains a timestamp, keywords, tags and context

  8. 8. Similarity proposes candidate links

    Vector search comes first; the model then decides which links to create

  9. 9. New evidence can revise an older note

    Memory evolves by updating context, keywords and tags

  10. 10. The answer receives selected notes

    The query embedding is compared with stored embeddings

  11. 11. The experiment tests dialogue question answering

    LoCoMo and DialSim; our focus is the GPT-4o-mini ablation

  12. 12. Links and updates help in this ablation

    GPT-4o-mini · LoCoMo · multi-hop F1, %

  13. 13. Useful organization still needs reliability checks

    Text QA supports the mechanism while leaving operational questions open

  14. 14. Storage and use are separate decisions

    RD-Forget retains an archive and builds a query-specific view

  15. 15. Disabling forgetting lowers scores in this test

    GPT-5.6-Luna · matched Full and −FORGET · accuracy, %

  16. 16. The next dependent task tests memory

    MemoryArena: action, feedback, update, another session

  17. 17. Partial progress does not guarantee task success

    Two Table 3 rows: long context and Mem0

  18. 18. In CI, measure outcomes and transfer

    Bridge to episode 42: isolate trials and keep the grader fixed

  19. 19. Forgetting needs a no-request control

    Bemula first checks that the system knew the target fact

  20. 20. At most 1.5% of requests fully pass

    RPR requires every request probe and every preservation probe to pass

  21. 21. Forgetting requests also affect derived records

    CI example: an internal address appears in a log, note and summary

  22. 22. Each memory operation needs its own contract

    Source, scope, revision and a checkable outcome

  23. 23. Useful memory changes a verifiable action

    From “the agent remembers” to understanding what it uses and why