Skip to slide
090 / 170

CH 09 · Memory Outside the Weights · 06/06

Read as one page

Key takeaways

  • The model is stateless; project knowledge lives outside the weights in files the harness loads.
  • CLAUDE.md / AGENTS.md are injected every session and reloaded after compaction. Put persistent rules there, not in chat.
  • Aggregate instructions "more specific wins": root first, working directory last; respect a size cap; load path-scoped rules only when relevant.
  • The two-tier pattern (shared rules in git, personal memory per user, with a promotion workflow) makes memory a reviewable team practice and keeps the whole setup portable.
  • Memory sits in the cached prefix, so a stable instruction file is both more effective and cheaper.

Original sources: Anthropic's How Claude Code works, the "Agentic Harness Architecture: Seven Pillars" breakdown, and OpenAI's Unrolling the Codex agent loop.

Finished this idea?

Your learning progress stays in this browser.