How to Organize What Your AI Knows: One Notebook per Project, Client or Person

Once an AI assistant can remember things, a new problem appears: it can remember too much, in the wrong place. A note about one client's pricing should never colour the advice you get about a different client. The fix is the same thing that has worked for filing cabinets for a century: separate folders.

Why one big memory backfires

  • Leakage. Details from one project show up in another where they do not belong.
  • Noise. The more unrelated notes the assistant sees, the less sharply it focuses.
  • Privacy. Sharing one memory with a collaborator shares everything in it.

A simple structure

Create one container per context you would naturally keep apart:

  1. Per project for goals, decisions, and open questions.
  2. Per client for preferences, history, and deliverables.
  3. Per person for yourself or a family member: routines, preferences, things to remember.
  4. One small "general" container for style preferences that apply everywhere.

What to store, and what to skip

Good candidates: decisions and the reason behind them, constraints, naming conventions, and "we already tried that" notes. Poor candidates: raw chat transcripts, anything that will be stale next week, and sensitive credentials.

Tools that follow this model

You can do this by hand with folders and a text file. If you want your assistants to read and write the notes themselves, look for memory services that model this explicitly. Coconut does: each "coconut" is a private notebook (internally its own knowledge graph) for one project, client or person, connected to assistants such as Claude, ChatGPT, Cursor, VS Code and Gemini CLI over MCP. A coconut can be shared by email with read-only or read-write access, and everything can be exported or deleted.

For code projects, separate "what we decided" from "what the code is"

Memory should hold your decisions. It should not try to replace an understanding of the code itself, which changes with every commit. For that, a structural index kept in sync with the repository, such as CodeMesh with its tree-sitter based graph, serves a different purpose and pairs well with a decision log.

Keeping it healthy

  • Review each notebook monthly and delete what is stale.
  • Date important decisions so conflicts are easy to resolve.
  • Prefer short, factual notes over long narratives.

Frequently asked questions

How many notebooks is too many?

If you cannot remember which one holds something, you have too many. Start with one per active project or client.

Can I share just one notebook with a colleague?

That is the main advantage of separate containers: sharing one does not expose the others.

Should the AI write to memory automatically?

Letting it save routine facts is convenient; review what it saved periodically so errors do not accumulate.

Read more