Production agent dev: every AI memory tool is a vector store in a trench coat
According_Bee_2957 · reddit · 2026-09-27
A developer building production agents complains that almost every memory tool is just a vector store with a wrapper: embed conversations, store them, and retrieve whatever looks most similar.
Real-world failure: a user said they moved from Delhi to Mumbai in session 4, and in session 7 the agent recommended a Delhi restaurant because the old embedding scored higher. When a preference changes, the store keeps both versions and the model picks randomly — none of which is a model problem, but the memory not knowing what changed.
What he wants: typed extraction (fact vs preference vs event), contradiction handling so old facts get replaced, and entity resolution so one person mentioned 10 ways stays a single entry. He's asking the community for options.
More from coding & agent
- Sydney Codex Community Buildathon set for Nov 14 with Cerebras-backed ultrafast compute track — paw_lean · 2026-11-11
- shadcn's improve hits 10k stars: let top models plan, cheap models execute code fixes — shadcn · 2026-11-11
- OpenAI open-sources Codex Security CLI and SDK for AI-driven vulnerability scanning and fixing — openai · 2026-11-11
- Claude Opus 5.5 runs an architecture search and finds AttnRes-Canon, a new top Transformer design — SonglinYang4 · 2026-11-11
- 5 research habits worth knowing: read papers backwards, reproduce first, wire coding agents via MCP — cneuralnetwork · 2026-11-11
- 'Adam can now do CAE': AI agent showcase in engineering simulation draws buzz — burhop · 2026-11-11