# Membase — full reference for AI readers > One memory. Every AI. Only yours. Membase is a long-term memory for you and your AI. Bring knowledge, preferences and decisions from conversations, files and connected apps into a lasting memory you and your AI can build on. You choose what to keep and who can access it. Membase is made by Unibase. Site: https://membase.com · Docs: https://docs.membase.com · App: https://app.membase.com Parent: https://unibase.com (docs at https://docs.unibase.com) --- ## What problem it solves Preferences, research and decisions build up over time, spread across conversations, files and apps. Membase brings them into a lasting memory you can update and put to use with your AI. --- ## The model Three nouns, and the same tool table behind every surface: - **Container** — a scope. One per project, client or part of your life. Access is granted per container, so an agent can be given a project without being given everything else. - **Memory** — a fact, a note, a saved answer. You can read, correct, delete and export any of them. - **Document** — a file or page you put in, which becomes retrievable alongside memories. Two verbs: - **`search`** returns passages, each with the container it came from. It is retrieval, not generation — nothing is synthesised. - **`ask`** returns an answer written over those passages, for when you want the agent to answer rather than to read. --- ## How to use it ### As a person - **Web app** — https://app.membase.com. Sign in and the memory is there. Nothing asks for a wallet. - **Chrome extension** — syncs conversations from ChatGPT, Claude, Gemini, Grok, DeepSeek and Kimi, and saves web pages. Encrypted on the device with a wallet-derived key before anything leaves it. - **Desktop app** — coming soon; local memory and offline search are planned. ### As a developer - **MCP** — point any MCP client at `https://api.app.membase.com/mcp-http`. One command connects Claude Code; Cursor, ChatGPT and other clients take the same URL. Sign in once; the consent screen asks which memories the client may reach. - **REST API** — the same tool table over HTTP. - **SDKs** — `pip install unibaseio-membase` (Python), `npm i @unibaseio/membase` (TypeScript). One client, one key. `pip install 'unibaseio-membase[local]'` adds the local engine, which is where unibaseio-membase-core comes in — a dependency of that extra, not something to install on its own. - **CLI** — `npx @unibaseio/membase-cli --help`, or install it globally. The binary is `membase`: `membase add`, `search`, `ask`, `documents add|list`, `profile`. A separate package from the SDK, because most people installing the SDK are not going to type at it. Tools available through MCP: `search_memories`, `ask_agent`, `add_memory`, `add_document`, `list_containers`, `get_profile`. --- ## What makes it different Membase is a product **and** an API over the same memory: 1. **You can see it.** Every fact it holds is readable, editable and deletable, and the whole thing is exportable in one request. 2. **Several agents share one memory.** Not one memory per app. Scoping decides which agent reaches what, and revoking takes effect at once. 3. **It reaches the assistants you already use**, through the extension, rather than only the agents you write. --- ## Benchmarks These measure the **memory engine**, not the product as a whole, and are reproducible from the published benchmark repository. | Benchmark | Accuracy | Mean context tokens | Against full context | |---|---|---|---| | LoCoMo | 93.1 | 6,562 | ~20k (3× less) | | LongMemEval | 92.6 | 8,970 | ~103k (11× less) | | DMR | 92.2 | 1,602 | — | --- ## Storage and privacy - The hosted service stores and processes the sources you provide to build and retrieve memory. Each account’s memory runs in an isolated environment. - The hosted service is not end-to-end encrypted. The extension has its own privacy policy describing its data handling: https://membase.com/privacy. - Unibase DA is available as a verifiable storage backend through the protocol SDK; the hosted product does not depend on it. --- ## Relationship to Unibase Unibase publishes an open protocol stack: **AIP** for ERC-8004 agent identity and cross-platform interoperability, **Unibase Pay** for x402 payments agents settle themselves, and **Unibase DA** for verifiable data storage. **UB** is the network token. Membase is a Unibase product and has its own site. It does not require any of the protocols to run. - Unibase: https://unibase.com - AIP: https://unibase.com/aip - Unibase Pay: https://unibase.com/pay - Unibase DA: https://unibase.com/da --- ## Common questions **What is Membase?** Membase is the memory layer for AI agents and the people who use them. It builds one long-term memory from your conversations, files and connected apps, and every AI tool you connect can read it, including ChatGPT, Claude, Cursor and your own agents. You choose what goes in and which tools can access it. Membase is built by Unibase. **Is Membase related to membase.so or the Membase database?** No. Membase (membase.com) is the AI memory layer built by Unibase. It is not affiliated with membase.so, and it is unrelated to the Membase key-value database that later became part of Couchbase. **Which AI tools work with Membase?** The Membase Chrome extension works with ChatGPT, Claude, Gemini, Grok, DeepSeek and Kimi. Coding tools such as Claude Code, Cursor and Codex connect through the Membase MCP server. Custom agents can use MCP, the REST API, the CLI, or the Python and TypeScript SDKs. **How do I add Membase to my agent or app?** Install the SDK with pip install unibaseio-membase or npm install @unibaseio/membase, set your API key, then add and search memories in a few lines of code. Any MCP client can also connect to the Membase MCP server directly. The quickstart is at docs.membase.com. **How accurate is Membase?** Membase answers 93.1% of LoCoMo, 92.6% of LongMemEval_S and 92.2% of DMR questions correctly. It sends the model about 67% fewer tokens than the full conversation history on LoCoMo, and about 91% fewer on LongMemEval_S. The full method and results are on the Membase benchmarks page: https://docs.membase.com/evaluation/benchmarks/ **How is Membase different from Mem0?** Both Membase and Mem0 give AI agents long-term memory through an API. Membase is built around one memory that you can use from consumer AI apps (through the Chrome extension) and from developer tools (through MCP and the SDKs). It runs hosted or locally, and its benchmark method and harness are published. **Can Membase run locally?** Yes. Membase runs hosted, or on your own machine with pip install 'unibaseio-membase[local]' (Python 3.12 or 3.13, with your own OpenAI API key). The same API and MCP tools work in both modes. **Is Membase open source?** The Membase SDKs, CLI and MCP server are public on GitHub at unibaseio/membase-ai. The core memory engine is proprietary and ships as a compiled package under the Unibase license. **Can I control what Membase remembers?** Yes. You choose what gets added: synced conversations, saved items, uploaded files, connected apps, or knowledge added by authorized agents. You can keep separate memories for different projects, review and correct what was saved, delete or export it, and revoke any tool’s access at any time. **How does Membase handle my data?** The hosted service stores and processes the sources you provide in order to build and retrieve your memory. Each account’s memory runs in an isolated environment. The hosted service is not end-to-end encrypted, so use local mode if your data must stay on your own machine. The Chrome extension has its own privacy policy.