Agent Memory Kit

The essential memory setup for the AI assistant you already work with, in three pieces: a settings block it loads every session, a project file it reads before anything else, and an optional skill that builds the structure and repairs it. Works in Claude Cowork, Claude Code, Codex and Cursor.

What it is

The memory layer of an AI operating system, and the piece almost nobody installs. Everything else you do with an assistant sits on top of it: without memory, you are re-briefing it every morning for the rest of your life, which is why most people quietly conclude their assistant is dumber than it is.

Right now it opens every session knowing nothing about your work: not what you decided yesterday, not who the client is, not why the last approach was dropped. The notes it does keep end up in a directory that gets wiped between sessions, or in one file while a second file that needed the same edit goes stale.

This kit is the protocol that closes that gap. A small set of rules your assistant reads before it touches anything, and a project file it keeps current while you work.

It comes in three pieces because they load at different moments, and merging them makes all three worse.

1. Global settings

A block you paste once per machine into your tool's global instruction field. It has to run before any file is read, which is exactly why it cannot live inside the project file. Under 2 KB, so it costs almost nothing per session. It defines when your agent reads memory (the map first, then the file whose subject the task actually touches) and when it writes.

2. Project template

A CLAUDE.md or AGENTS.md for the project root, carrying the three sections the global block checks for: a memory map, an active tasks table, and update rules. One per project. This part is the content, not the protocol.

3. Skill

memory-structure, optional. It holds the spec for the instruction file and the logic for repairing one that is missing a section, and it loads only in sessions where the structure is actually being built or fixed, which for most projects is one session in twenty.

Where each piece goes

ToolGlobal settingsProject file
Claude CoworkSettings, Cowork instructionsCLAUDE.md
Claude Code~/.claude/CLAUDE.mdCLAUDE.md
Codex~/.codex/AGENTS.mdAGENTS.md
CursorSettings, Rules, User RulesAGENTS.md

Piece 1 goes in once per machine, piece 2 once per project. The skill goes to your tool's skills folder when it has one, and gets pasted into the chat on the day you need it when it does not.

The rules your agent ends up following

  • Read the project instruction file before anything else, check that its three sections are there, and offer to fix the structure instead of quietly rewriting it.
  • Never create an instruction file or a memory folder unasked. The folder appears at the moment of the first real write, and your agent says so in that turn.
  • Write at four moments, without being told: end of session, a decision that changes how future work happens, a new person, client, project or term, and anything already in memory that turned out to be wrong.
  • Never record an inference as a fact. Anything unconfirmed is written as an assumption or an open question, and labelled as one.
  • Never write credentials, keys or tokens into a memory file.
  • Never write memory outside the project, and never into a temp or session directory, because some tools wipe those between sessions.
  • Show the diff before saving, and replace a superseded fact rather than stacking the new version underneath the old one.

    • Installing it

    Hand the whole folder to your agent and say "install this kit". It reads 0-INSTALL-ME.md, states which tool it is running in, prints the block you have to paste yourself wherever it cannot write to settings, sets up one project, and finishes with a list of what it wrote and what it skipped. It asks before every write, and it never overwrites an instruction file you already have. About five minutes, most of that reading.

    What to check after a week

    Open the memory folder and read what is in it. If guesses are sitting there written as facts, the protocol is not doing its job and the file needs pruning. That failure mode is a quiet one, and it is the one worth watching for.

    Take it further

    The kit gives you the protocol, and the protocol has a ceiling: it remembers what you or your assistant remembered to write down, in the one project you set up. Good sessions still slip through on the days you are moving fast, which are exactly the days worth keeping.

    The Memory Engine takes the ceiling off. Every session is captured on its own, with no one deciding what deserves saving. Recall runs by search across everything you have ever worked on, instead of by opening the one file whose name you happened to remember. Your assistant walks into a brand new project already holding the decisions, the people and the constraints from all the others, on day one.

    That is the difference between an assistant that keeps notes and one that has a memory.

    See how the Memory Engine works