Claude memory system

How Claude remembers your work

Every request leaves a note in the project’s memory folder. Sync copies it between machines, and lookups read it back.

Step Check Blocked File saved

The full workflow

One request from start to finish: who acts, where each step saves, and where it ends up later.

Every turn

From your request to a saved note. Each row shows who is acting.

Sync between machines

Run /memory-sync when you switch machines. Each row is a place the notes pass through.

What’s in memory/

Generated files stay small. One lookup script searches the notes and every history entry; grep is only a fallback.

INDEX.mdOrientation and how-tos, no note list Read first, before changing code TAGS.mdThe allowed tags and their meaning <folder>/<type>-<name>.mdEach note, with a one-line Summary: history/YYYY-Www.mdOne entry per prompt, one file per week Weeks start on Sunday. Searched by the lookup script The ID names its file: H-2026W42-0042→history/2026-W42.md build-refs.shevery turn REFS.mdgenerated Notes, tags, and a list of months Read in full. About 20 KB after a year refs/YYYY-MM.mdgenerated Entries with a Did, plus one line per week Read in full. About 40–50 KB a month month list Written by Claude or hooks Generated, never edited by hand

Finding past work

What Claude does when you ask about something you did before. The lookup script always runs first.