Skip to main content
Glama
README.md
<p align="center"><img src="design/logo/export/ownmemory-logo-light.png" alt="OwnMemory" width="440"></p>

OwnMemory lets Claude remember you across chats and devices, with the memories stored encrypted under a key only you hold, in storage you own.

Claude calls a remote MCP server running in your own Cloudflare account. The server encrypts each memory and stores it in a D1 database in that account (the one-click default) or in an `OwnMemory` folder in your Google Drive. It works in the official Claude apps on the web, iPhone and Android.

[![Deploy to Cloudflare](https://deploy.workers.cloudflare.com/button)](https://deploy.workers.cloudflare.com/?url=https://github.com/ownmemoryhq/ownmemory/tree/main/server)

**Get started:** create a free Cloudflare account and connect GitHub to it first ([steps](docs/setup.md#one-click-setup)), then click the button above, choose a random owner code of 12+ characters, and open your server's address to follow its four setup steps. About 5 minutes, on a free Cloudflare account. Full guide, including Google Drive storage: [docs/setup.md](docs/setup.md).

**Status:** early (v0.2). Built and tested for Claude: on the web, iPhone and Android, and in Claude Code. Other MCP clients that support remote servers with OAuth should be able to connect, but are untested.

## How your memories stay yours

![Key flow](docs/images/key-flow.png)

- **Your passphrase never leaves your browser.** When you connect Claude, the passphrase page turns it into a key that unlocks a random 256-bit memory key. Only the locked copy of that key is stored, in your storage (`keyinfo.json`).
- **Every memory is encrypted** (AES-256-GCM) before it is stored, in a file with a random name. Your storage provider and anyone with the files see only ciphertext.
- **The server keeps no key at rest.** The memory key and your Google token live only inside the OAuth grant, which is encrypted with Claude's access token. The MCP handler is stateless: nothing is written to Durable Objects or logs.
- **Hard rules in code.** Blocked categories, a pause switch, and a filter that refuses passwords, API keys and card numbers, whatever the AI decides.
- **Search by meaning, still private.** Embeddings come from Cloudflare Workers AI in your own account and are stored encrypted like the memories. Set the `SEMANTIC_SEARCH` secret to `off` to keep memory text away from it; search then uses keywords only.
- **See and fix what it remembers.** Your server's `/memories` page lets you browse, search, edit, add and delete memories, decrypted only in your browser.
- **No lock-in.** **Download all** on the memories page gives you every memory as Markdown, and [ownmemory.app/decrypt](https://ownmemory.app/decrypt) decrypts a downloaded copy in your browser, even without your server.

The one unavoidable exposure: the server and Anthropic see a memory's text while a save or search runs, because Claude sends plain text. Running the server yourself is the safeguard.

Who else could get at your memories, person by person, and how to keep them out: [docs/security-model.md](docs/security-model.md).

## Tools Claude gets

| Tool | What it does |
| --- | --- |
| `load_core_memories` | Called at the start of a chat: returns the memory mode, any message for you, and (in OwnMemory-only mode) your profile, preferences and recent work |
| `search_memories` | Search by meaning and keywords, English, Chinese or mixed (finds "What do I like to drink?" without shared words) |
| `list_memories` | Most recent memories, optionally by category |
| `save_memory` | Saves one fact, after the hard rules allow it; points out facts already saved, even when worded differently |
| `import_memories` | Saves up to 50 facts at once, for example everything Claude's built-in memory knows |
| `update_memory` | Changes a fact, keeping the previous wording as dated history |
| `delete_memory` | Deletes a memory; it stays 30 days in D1, or in the Drive trash |
| `review_memories` | Finds memories with the same wording or meaning, to merge |
| `set_memory_mode` | Records how you use OwnMemory alongside Claude's memory: ramp-up, OwnMemory only, or keep both |

## Repository layout

| Path | What's there |
| --- | --- |
| `server/src/` | The Cloudflare Worker: OAuth and sign-in flow (`auth.ts`), passkey and authenticator sign-in on one-click servers (`gate.ts`, `factors.ts`), the memories page's API (`memories.ts`), MCP tools (`tools.ts`), storage backends D1 and Drive (`storage.ts`, `google.ts`), encrypted memory store (`store.ts`), memory format, search and duplicate checks (`format.ts`), memory modes and ramp-up (`settings.ts`), pages (`pages.ts`) |
| `server/public/unlock.js`, `home.js`, `memories.js`, `gate.js` | Scripts for the passphrase page, the setup wizard, the memories page and the passkey / authenticator sign-in step |
| `server/public/keys.js`, `memory-file.js`, `zip.js` | Encryption, the memory file format and the zip builder, shared by the browser, the Worker, the decrypt tool and `site/public/decrypt.html` (copies; `npm run sync-site`), so the format can't drift |
| `site/` | The ownmemory.app landing page and [the decrypt page](https://ownmemory.app/decrypt): static files, no data collected |
| `server/scripts/decrypt.mjs` | Local export to plain Markdown |
| `server/test/` | Unit tests (`npm test`) and a recall evaluation against real embeddings (`npm run eval`) |
| `docs/` | Setup guide, "Who can read your memories" and the key-flow diagram |
| `design/logo/` | Logo sources from the design canvas and PNG/SVG exports |

## Contributing, security and license

- Contributions are welcome; see [CONTRIBUTING.md](CONTRIBUTING.md). Contributors sign a short [contributor license agreement](CLA.md).
- Found a security problem? Please report it privately: [SECURITY.md](SECURITY.md).
- Licensed under the [Apache License 2.0](LICENSE). "OwnMemory" and the OwnMemory logo are trademarks of the project and are not covered by the license; forks need a different name.

Maintenance

ActivityMaintained
ResponsivenessNo issues