Skip to content

Agent memory layer for cross-run learning #231

Description

@evansenter

Summary

Agents are stateless. A DeepResearchAgent that discovers a useful decomposition pattern doesn't share it with future runs. Knowledge is lost at session end.

Proposal

Add a memory layer that persists useful patterns across agent runs:

  • Pattern storage (successful strategies, useful decompositions)
  • Retrieval at agent start (seed new runs with past learnings)
  • Decay/relevance scoring (old patterns may become stale)

Open Questions

  • Where to store? (SQLite, filesystem, MCP resource?)
  • What to store? (Explicit agent outputs? Implicit behavioral patterns?)
  • How to retrieve? (Semantic search? Structured queries?)
  • How to avoid context bloat? (Summarization? Relevance filtering?)

Context

From user_guide.md "What to Build Next" - Priority 2.

Suggested by gemicro session during cross-session Frontier feedback.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions