Stop Building Agent Memory Like Chat Logs
Cold-Cranberry4280 · reddit · 2026-07-16
Drawing from a year of experience with agent memory, the author argues that "storing everything and doing RAG" is a fundamentally flawed default approach.
The core issue isn't just retrieval quality, but the lack of structure in raw transcripts:
- Outdated facts and new facts remain equally retrievable.
- Lack of coreference resolution makes it impossible to track who "he," "that client," or "this project" refers to.
- The relationship between related messages regarding the "same event" is lost.
The author advocates treating memory as a "user's world model" rather than a chat log:
- Treating entities like people, projects, and commitments as first-class objects.
- Attaching source, confidence, and timestamps to every fact.
- Directly updating the structure with new info instead of piling it alongside old facts.
- Retrieving from a structure that understands internal relationships, rather than relying on text similarity.
This approach acts as a lightweight temporal knowledge graph, though the author admits challenges remain in deciding when to overwrite old facts versus preserving historical changes.
More from coding & agent
- A Forward Deployed Engineer job really has three stages: audit, evals, deploy — blaizedsouza · 2026-07-22
- 438 sealed tests show coding agents prefer DIY over third-party databases — cramforce · 2026-07-22
- Building a stock research agent turned out to be a prompt-design problem, not a data problem — jwstrategies · 2026-07-22
- Tenable and AWS launch a Black Hat build event for open-source security agents and MCP servers — Dave_Maynor · 2026-07-22
- Discussion: How Much Permission Would You Give Your Background AI Agents? — Total_Drag7439 · 2026-07-22
- Codex helps build Valdiluce, an open-world game with climbing, gliding and gondolas — Dimillian · 2026-07-22