# Cozy compared with built-in AI assistant memory

> How Cozy (cozy.cloud) differs from ChatGPT Memory and Projects, Claude memory and Projects, and Gemini memory: shared wiki, hosted MCP, scoped access, and where built-ins win.

Updated 2026-11-09. Canonical: https://cozy.cloud/compare/built-in-ai-memory

> Built-in memory in ChatGPT, Claude and Gemini is the easiest way to get an assistant to remember you, but each store belongs to one vendor. Cozy (cozy.cloud) is a hosted Markdown wiki that any MCP-capable agent can read and write, so one memory is shared across assistants and you can inspect, edit and restore it.

This page compares the Cozy wiki at cozy.cloud with the memory and project features inside ChatGPT, Claude and Gemini. Cozy is operated by Coral Sourcing LLC and is not affiliated with Cozy Cloud (cozy.io) or Cozi. We make Cozy, so read the claims with that in mind. We have tried to say plainly where the built-in features are the better choice.

## What built-in memory does well

Built-in memory needs no setup. You turn it on, talk to the assistant, and it picks up preferences and facts over time. Projects add a place to keep instructions and files for a topic. For one assistant used by one person, this is often enough, and Cozy adds nothing you need.

Anthropic made memory and a memory import tool available on Claude's free plan on 2 March 2026, according to [MacRumors](https://macrumors.com/2026/03/02/anthropic-memory-import-tool/). The import is a paste-in of another chatbot's output, so it is a one-time snapshot and not continuous sync.

We did not verify Gemini's memory behavior against Google documentation for this page, so we make no specific claims about it. Check Google's current help pages before deciding.

## Where the approach differs

Cozy runs no model. You bring your own agent, and the agent reaches your wiki through a hosted MCP endpoint at `https://cozy.cloud/mcp`, one URL for every workspace in your account. Claude, ChatGPT clients that support MCP connectors, Codex, Cursor and Claude Code can all point at the same URL. See [the MCP server page](/mcp-server) for the transport and tool list.

Built-in memory is different by design. Per [MemoryPlugin's comparison](https://blog.memoryplugin.com/claude-vs-chatgpt-vs-gemini-memory/), updated 28 July 2026, ChatGPT memories cannot be imported from other assistants and are not documented as part of the standard export. The same source says memory holds high-level context rather than verbatim text. In Cozy, memory is a set of Markdown pages you can open, diff and edit.

## Side-by-side

| Question | Built-in memory | Cozy |
|---|---|---|
| Who can use it | The one vendor's assistant | Any agent that speaks MCP |
| Format | Opaque to you, vendor controlled | Markdown, text, JSON, CSV files |
| Editing | Through the assistant UI | Edit pages directly or via agent tools |
| History | Vendor dependent | Last 50 revisions readable and restorable |
| Provenance | Not documented as cited | Links into sources/ become cites edges |
| Access control | Account level | Scopes read, wiki-write, source-write plus folder prefixes |
| Setup effort | None | Create a workspace, connect an agent |
| Model included | Yes | No, bring your own |

## Projects versus workspaces

ChatGPT Projects and Claude Projects group chats, instructions and files. Cozy has a similar idea in workspaces: every account gets "My wiki", and you can create more private workspaces to keep projects apart. The difference is that a Cozy workspace is addressed by one URL that several assistants share, and its pages carry typed links, tags and status that Cozy derives from Markdown conventions on every save.

If you only want a pile of files attached to one assistant's chats, a built-in Project is simpler. If you want the same notes to show up in Claude at your desk and ChatGPT on the web, a Cozy workspace fits better.

## Control and safety

Cozy agents authenticate with OAuth 2.1 (S256 PKCE, refresh tokens), capped by the scopes granted and the person's own access, and can be disconnected under Connectors → Agents. Scripts can use bearer credentials created in the Agents screen. Credentials are random, shown once, stored as SHA-256 digests, revocable, and expire within 90 days. Scopes never exceed your own access, and folder-prefix grants can limit an agent to part of the wiki. Agents cannot see connector credentials or permanently purge.

Edits require `expectedRevision`, so concurrent writes return a conflict rather than overwriting. Deleting a page writes a tombstone, and moves keep history.

## Keeping memory healthy

Cozy checks the wiki for broken links, duplicate titles, stale pages whose cited source changed, conflicts, orphans, unsourced pages and oversized pages, then returns a score and a prioritized list of fixes. An optional Dreaming routine, run by your own agent, consolidates recent changes, records `CONFLICT:` lines and merges duplicates. Read more on [wiki health](/wiki-health) and [memory consolidation](/memory-consolidation).

## Where Cozy is the weaker choice

- Cozy has no built-in model. You need an agent and a client that supports MCP connectors.
- Search is lexical. There are no embeddings or vector database.
- Built-in memory learns automatically from conversation. Cozy stores what you or your agents write, so someone has to ask the agent to save things.
- Third-party sources report that write-capable custom MCP connectors in ChatGPT are limited to certain plans, so check [OpenAI's current documentation](/connect/chatgpt) for your plan.

## When to use both

Many people will keep built-in memory switched on for casual preferences and use Cozy for durable material: project notes, decisions, meeting evidence and source documents. For a broader look at the category, see [AI memory](/ai-memory), and for setup see [connecting Claude](/connect/claude).

## Frequently asked questions

### Is Cozy a replacement for ChatGPT Memory or Claude memory?

No. Cozy runs no model, so it complements an assistant. You connect your own agent to a Cozy workspace over MCP and keep built-in memory on or off as you prefer.

### Can Claude and ChatGPT share one Cozy memory?

Both can be pointed at the same URL, https://cozy.cloud/mcp, if the client supports MCP connectors. Write access in a given ChatGPT plan may be restricted, so check OpenAI's documentation. See [connect ChatGPT](/connect/chatgpt).

### Can I export my memory from Cozy?

Cozy files are Markdown, text, JSON or CSV, so the content is plain and portable. Deleted pages are tombstoned and the last 50 revisions stay readable and restorable.

### What does Cozy cost?

Cozy has no public pricing, so this page makes no price comparison. Built-in assistant memory is bundled with the assistant plans you already use.
