# LiteLLM Memory
See what your team is working on. Ask about anything shared with you.
A [LiteLLM Labs](https://github.com/BerriAI) product, built with [Ente](https://github.com/ente-io/ente) and [LiteLLM](https://github.com/BerriAI/litellm).
LiteLLM Memory records the work people do with AI and makes it easy to ask about later. Connect Codex or Claude Code, browse your company’s shared work and your own private history, and ask questions in the web app or through **@Memory** in Slack.
This is a work archive you query deliberately. It does not search old memories or inject context into every coding turn.
> **Release status:** prerelease. Review the [release checklist](docs/release-checklist.md) before using real team history.
The actual application, shown with synthetic local records. See [Ask](docs/assets/ask.jpg), [privacy settings](docs/assets/settings.jpg), and the [interface guide](docs/ui-review.md) for more screenshots and verification details.
## What it does
- **Browse:** Company opens first, with a chronological feed and Slack profiles. Open a teammate’s shared activity or switch to My work.
- **Daily summaries:** browse a day of company work or your own history, or switch between Activity and Daily summaries on a person’s page. Complete days generate automatically from all of that day’s accessible memories, in chronological order.
- **Record:** connect Codex or Claude Code once. Your agent saves useful work context and can skip turns with nothing new. Connection status shows when work actually arrives.
- **Share deliberately:** work starts private. Preview a memory before sharing it, or opt in to sharing new memories. Agent-written memories are kept as written, without another model rewrite. TypeSafe Jev checks your private topics in parallel and caches completed checks.
- **Ask:** get cited answers through the web, Slack, API or MCP, limited to the evidence the audience can access. Daily summaries are available in the app and to the same agent in Slack. No automatic recall on coding turns.
- **Use your gateway:** model choices come from LiteLLM. Optional [Pydantic Logfire](docs/observability.md) shows usage and latency without private content.
See the [member guide](docs/member-guide.md) for filtering, privacy, exports and account controls.
## Deploy
[](https://render.com/deploy?repo=https://github.com/BerriAI/litellm-memory/tree/main)
The Blueprint provisions the web app, a persistent volume and PostgreSQL. Ente Museum runs privately alongside the app and is provisioned automatically. Review Render's charges before creating the deployment; durable hosting is not a free-tier promise. Fork and deploy it with your own infrastructure and gateway.
**On any Docker host:**
```sh
git clone https://github.com/BerriAI/litellm-memory.git
cd litellm-memory
./scripts/self-host.sh
```
Open the address shown by your deployment. Setup stays on one page:
1. Claim the installation using its deployment setup code.
2. Connect your **LiteLLM URL and API key**. LiteLLM Memory discovers models, suggests defaults, and checks Jev through the gateway.
3. Connect Slack and finish with your verified company identity. Slack returns you to the same setup page.
Members then use **Sign in with Slack**. They do not handle encryption keys, recovery phrases or a separate LiteLLM Memory password. Eligibility is restricted to the configured Slack workspace and exact verified email domain. Choose your company domain during setup.
**Jev is required.** Built-in support starts in LiteLLM **v1.103.0-rc**. Use a gateway release with the [TypeSafe integration](https://docs.litellm.ai/docs/pass_through/typesafe), and configure `TYPESAFE_API_KEY` on that gateway. Setup explains missing capabilities and retains your inputs. LiteLLM Memory never asks members for a TypeSafe key or bypasses the gateway. See [gateway setup](docs/self-hosting.md#connect-typesafe-jev-through-litellm).
Read [self-hosting](docs/self-hosting.md) for HTTPS, backups, updates, external Ente and configuration.
## Connect your tools
Open **Connections → Connect a tool** in LiteLLM Memory. On the computer running your coding tools:
```sh
uv tool install --force 'git+https://github.com/BerriAI/litellm-memory.git@main'
litellm-memory-connector connect https://your-litellm-memory.example
```
Sign in with Slack and approve the connection and reminder hooks in your browser. One setup connects Codex and Claude Code using the same encrypted sign-in, while preserving unrelated settings. The browser returns to a first-save guide. Reopen your coding tools, send the provided test message in a new task, and confirm the memory appears in My work. Use `--host codex` or `--host claude` to configure only one.
The connector reminds your coding agent to save useful work memories through MCP: current work, progress, findings, decisions, blockers and next steps. It does not automatically upload conversations. The agent can skip a turn when nothing useful changed. Optional global agent instructions are available in the [connector guide](docs/connectors.md). The website and Slack work without installing the connector.
[Member guide](docs/member-guide.md) · [Connector guide](docs/connectors.md) · [API and data formats](docs/api.md)
## How privacy works
LiteLLM Memory encrypts persisted records with Ente-compatible libsodium encryption and stores ciphertext in Ente Museum. Each member has a separate wrapped key. Original evidence, derived records, public summaries, profiles, jobs and configuration are encrypted at rest; search indexes live in bounded process memory.
**The hosted application is a trusted processor.** It can decrypt authorized work to search it and call the LiteLLM gateway. A person controlling the runtime and its keys could also do that. This is not Ente Photos' zero-knowledge client model, and we do not claim protection against a malicious deployment operator.
Normal administrator interfaces cannot inspect someone else's private records. API authorization is checked before decryption and again before delivering agent output. Publishing shares a separate summary, while the original evidence stays private. Requests ask LiteLLM to suppress content logging, and agent tracing is disabled; the gateway and model provider still need appropriate retention settings.
Read the [security and privacy model](docs/security.md) before operating an installation.
## Develop
```sh
uv sync --frozen
uv run pytest -q
uv run ruff check .
```
The app is `litellm_memory.application:app`; the portable launcher is `python -m litellm_memory.boot`.
[Architecture](docs/architecture.md) · [Product specification](docs/product-implementation-spec.md) · [Release checklist](docs/release-checklist.md) · [Answer-quality evaluation](docs/answer-quality.md)
See [CONTRIBUTING.md](CONTRIBUTING.md) for local development, code orientation and review expectations.
## License
LiteLLM Memory is licensed under **AGPL-3.0**. The bundled Ente Museum executable is unmodified and also AGPL-3.0; see [LICENSE](LICENSE) and [third-party notices](THIRD_PARTY_NOTICES.md). The self-hosted core needs no proprietary LiteLLM Memory control plane. Hosting and model inference may have costs.
Code belongs here; personal histories, credentials, deployment state and database exports do not. Keep those out of Git, even in a private fork.