How to stop re-explaining context to ChatGPT and Claude

Updated · By the Cozy team

To stop re-explaining context to ChatGPT or Claude, move it out of the chat and into files that outlive the session. Write project facts, decisions and preferences as notes in one shared place, then give every agent read and write access to that place. A new session reads the notes first instead of asking you again.

Why the AI forgets your project

A chat session holds context only while it is open and only up to the limits of the model's window. When you start a new session, switch from Claude to ChatGPT, or move from your editor to a terminal agent, the background you explained is gone. Some products offer built-in memory, but what it keeps is chosen by the vendor, tied to one product, and hard to inspect or correct.

The result is the same loop every time. You paste the project summary, restate your naming rules, repeat last week's decision, and correct the same wrong assumption again. The context exists, but it lives in your head and in old chat logs.

Where an agent should store notes

An agent needs a place to put notes that meets four conditions:

  1. Outside any one chat or vendor. Notes must survive a new session and be readable by a different agent.
  2. Plain and inspectable. Markdown files you can open, edit and diff are easier to trust than hidden memory.
  3. Reachable by tool call. The agent should read and write through a standard protocol, not through copy and paste.
  4. Safe to write to. Two agents editing at once should not silently overwrite each other, and a bad edit should be reversible.

A local folder of Markdown meets the first two. It often fails the third when you use a web or phone client, and the fourth when several agents write at once. A hosted wiki with an MCP endpoint covers all four. The idea is described in more detail on the LLM wiki page.

How Cozy handles it

Cozy is a private wiki and cloud file space for each person and their AI agents. Every account gets "My wiki", and you can create more private workspaces to keep projects apart. Cozy runs no model itself. You bring your own agent, and the agent reads and writes the wiki through a hosted MCP endpoint, described on the MCP server page.

Your account has one MCP URL for all your workspaces:

https://cozy.cloud/mcp

Tools use My wiki unless the agent passes a workspace argument, and cozy_workspaces lists your workspaces.

The wiki is organized into four folders:

The handoff folder is the direct answer to re-explaining. At the end of a session, an agent writes what it learned and what is unfinished. The next session, from any agent, reads that note first.

A practical routine

You do not need a large setup. Start small and let the wiki grow from real work.

  1. Create a project page. Write the goal, the constraints, the names you use, and the decisions already made. Keep it short.
  2. Connect your agent. Add https://cozy.cloud/mcp to the client you use and sign in. The Claude Code, Codex and ChatGPT pages cover the details.
  3. Tell the agent to read before it works. A line in your agent instructions such as "read the project page with cozy_context before starting" is enough. The cozy_context tool returns a page with typed neighbors, excerpts, evidence and warnings.
  4. Ask it to write back. Have the agent use cozy_append or cozy_edit to record new decisions and open questions, and to leave a handoff note under agents/.
  5. Review changes. cozy_changes lists what changed since a timestamp, so you can skim what the agent wrote.

For Claude Code, this is the whole setup, followed by signing in with /mcp:

claude mcp add --transport http --scope user cozy https://cozy.cloud/mcp

For Codex, run codex mcp add cozy --url https://cozy.cloud/mcp, which opens a browser to sign in. In claude.ai or ChatGPT, add a custom connector with the same URL.

What a stored note looks like

Cozy builds a graph from Markdown conventions every time a page is saved. A page can carry frontmatter for type, description, aliases, tags, status and as_of. Links between pages become typed relations, so a decision page can point to the project it belongs to. The typed links page explains the syntax.

Because pages are ordinary Markdown with wikilinks, they stay readable outside Cozy. The agent finds them with cozy_search, which is lexical and paginated, or by walking neighbors with cozy_graph. There are no embeddings and no vector database, so a search matches the words you and your agent actually wrote.

Keeping notes trustworthy

Stored context goes stale. Cozy has several guards against that.

Access and limits to know about

Agents sign in with OAuth 2.1, and you approve access on Cozy's consent page. An agent gets no more access than you have in each workspace, capped by the scopes it was granted (wiki:read, wiki:write, sources:write), and you can disconnect it under Connectors → Agents. For scripts and headless agents, each workspace also has an endpoint at https://cozy.cloud/wiki-mcp/WORKSPACE_UUID that takes a bearer credential you create in Agents. Credentials are shown once, stored as SHA-256 digests, can be revoked, and expire within 90 days. Scopes are read, wiki-write and source-write, with optional folder-prefix grants, and they never exceed your own access. Agents cannot see connector credentials or permanently purge.

Curated files can be up to 90,000 bytes. Imported sources can be up to 1,500,000 bytes. The search is lexical only, so it will not find a note by meaning when none of the words match. Write clear titles and aliases to compensate.

When something else is enough

If you only work in one product and its built-in memory holds what you need, you may not need anything more. If your notes live in a local folder and you only use a desktop agent on that machine, a local Markdown setup can be simpler. Cozy is aimed at the case where several agents, several devices or several people need the same notes. See the comparison with built-in AI memory for a longer treatment.

Frequently asked questions

Why does my AI forget my project between chats?

A chat holds context only for the open session and the model's window. A new session starts without it unless something supplies it again, such as pasted text, built-in memory or notes the agent reads through a tool.

Where should my agent store notes?

In plain files outside any single chat, reachable by a standard protocol such as MCP, with revision checks and history. In Cozy that means a workspace wiki with folders for pages, sources, skills and agent handoffs. See MCP server.

Does Cozy include its own AI model?

No. Cozy runs no model itself. You bring your own agent, such as Claude, Codex or ChatGPT, and it reads and writes the wiki through the hosted MCP endpoint.

Will two agents overwrite each other's notes?

Edits require an expectedRevision, so a concurrent edit returns a conflict instead of silently replacing the other change. Earlier revisions can be restored, as covered in multi-agent memory.

Give your agents a memory worth keeping

Cozy gives every account a private wiki that Claude, Codex, ChatGPT and other MCP clients can read and maintain.

Start your wiki