Skip to content

memory.md

Latest commit

 

History

History
58 lines (43 loc) · 2.46 KB

memory.md

File metadata and controls

58 lines (43 loc) · 2.46 KB
Raw

Memory files

What it is: Markdown files that onboard your agents. They're loaded into the agent's context so it knows your preferences (global) and how a given project works (project-level). This is how you "ramp up your crew" in the video — no fancy memory system needed, just markdown that improves over time.

Two levels

Level File Loaded Size
Global ~/.claude/CLAUDE.md (+ AGENTS.md) every session, every project keep tiny — it costs tokens on every request
Project <repo>/CLAUDE.md (+ AGENTS.md) sessions in that repo can be verbose; it's the repo's collected learning

The symlink trick (one file, all agents)

Claude Code reads CLAUDE.md; Codex / opencode / most others read AGENTS.md. Instead of maintaining two copies, point them at one file.

Our setup goes one step further: the source of truth lives in this repo (config/memory/global.md) and the agent paths symlink to it — so your global memory is version-controlled and travels to a new laptop.

./setup.sh memory
# links ~/.claude/CLAUDE.md, ~/.codex/AGENTS.md, ~/.config/opencode/AGENTS.md
#   -> config/memory/global.md

Edit one file (config/memory/global.md), every agent sees the change.

Per-project memory

cp <repo>/path/to/project-template.md  <yourproject>/CLAUDE.md
cd <yourproject> && ln -s CLAUDE.md AGENTS.md   # both agents share it
git add CLAUDE.md AGENTS.md

Template: config/memory/project-template.md.

How to build good memory

  • Global (config/memory/global.md): only durable personal preferences. Starter rules from the video: no em-dashes; don't over-weight dev cost; reproduce bugs end-to-end first; prefer e2e tests over unit tests.
  • Project: don't write it all by hand. When an agent gets something wrong, correct it and say "remember this in the project memory file so you don't repeat it." Over many sessions the crew gets smarter.
  • Trim with skills: move conditionally-needed instructions (e.g. "how to run e2e tests") out of memory into a skill so they only load when relevant.

Gotchas

  • Global memory bloat = silent token tax. Kun keeps his ~27 lines. If it grows, move things into skills or project memory.
  • The symlink is tracked by git, but its target (this repo path) must exist on the machine. Running ./setup.sh memory recreates the links on a new laptop.