Memory

memory — durable knowledge in markdown

What this is

A portable long-term memory: a memory/ folder where each fact is one small file, indexed by a MEMORY.md you load at session start. No runtime — just files any agent can read and write.

Layout

memory/
├── MEMORY.md              ← index: one line per fact, loaded every session
└── <slug>.md              ← one fact per file, with frontmatter

Each fact file:

---
name: <short-kebab-slug>
description: <one line — used to judge relevance on recall>
type: user | preference | project | reference
---
<the fact. link related facts with [[other-slug]].>

MEMORY.md line: - [Title](slug.md) — short hook. One line each. Never put fact bodies in the index.

What to save

• user — who they are: role, stack, expertise, hard preferences.
• preference — how they want you to work (a correction, a confirmed approach) + why.
• project — goals/constraints not derivable from the code or git history; convert relative dates to absolute.
• reference — pointers to external resources (URLs, dashboards, tickets).

What NOT to save

• Anything the repo already records (code structure, past fixes, git history).
• Things that only matter to the current conversation.
• If asked to remember something obvious, ask what was non-obvious about it and save that.

Recall

At session start (or when a task touches a known area), read MEMORY.md; pull the fact files whose description looks relevant. Treat recalled facts as point-in-time — if one names a file/flag/function, verify it still exists before relying on it.

Hygiene

• Before saving, check for an existing file that covers it — update, don't duplicate.
• Delete facts that turn out wrong.
• Link related facts with [[slug]]; a link to a not-yet-written slug is fine.

Notes

Pure markdown, no deps — works in any agent. Pairs with handoff (session-state) and corp (give agents a shared STATE.md).