Nate Berkopec: agent memory for preferences and knowledge is fine, procedures are not
nateberkopec · x · 2026-11-16
Nate Berkopec responds to the coding-agent memory debate with a taxonomy:
- Memorizing personal preferences is fine;
- Memorizing knowledge is good;
- Memorizing procedures is probably bad — procedures belong in the codebase to minimize drift, and agents shouldn't reprogram the factory without human approval.
He previously said there is surely a place for memory/LLM-wikis, but the default "hoover everything in and spit it out everywhere" approach is awful.
Related event: Debate over coding agent memory: Sentry founder vs Rails expert(2 posts)→
More from coding & agent
- Humble Bundle update removes 2FA and breaks core features; site down to two engineers — jonerp · 2026-11-16
- Wrap: an AI-first Mac recorder that logs what you did, not just what you said — therealdanvega · 2026-11-16
- Sentry CEO: agent 'memory' is really just knowledge graphs — zeeg · 2026-11-16
- Agent Water Cooler: an MCP server where agents swap real-world task playbooks — Boring_Presence_425 · 2026-11-16
- Sentry founder David Cramer defends agent memory: Junior extracts knowledge from tool responses — zeeg · 2026-11-16
- Hooking GSC + GA into an SEO agent: automated search-data diagnosis in practice — yangyi · 2026-11-16