Membase

[Membase]

Membase

One memory. Every AI. Only yours.
Teach it once. Every AI remembers.

[PyPI] [npm] [MIT]

Website · Docs · Web app · Discord · X

[Add sources, manage one memory, use it from every AI and app]

Membase is the memory layer for AI agents. It keeps one memory for everything you and your agents know, and every AI you use reads it. You stop re-explaining yourself each time you switch models or tools.

• Add. Facts, documents, web pages and chat exports from Claude or ChatGPT go into one memory.
• Manage. Memories keep projects apart; search them, ask them questions, forget what should go.
• Use. Your code, Claude Code, Cursor, Codex and ChatGPT read the same memory through the SDK or MCP.

This repository is the developer side: Python and TypeScript clients, the membase command line and an MCP server. The memory lives in your Membase account, which the web app and browser extension also fill, or on your machine with unibaseio-membase[local].

Write once, read from any AI

[Your app writes a memory; any AI reads it, here Claude Code]

Recordings: Python SDK (57 s) · Claude Code (1 min 33 s). The answer above is illustrative.

Accuracy

[LoCoMo 93.1%, LongMemEval_S 92.6%, DMR 92.2%]

93.1% on LoCoMo, 92.6% on LongMemEval_S and 92.2% on DMR, using about 3× fewer tokens than the full history on LoCoMo and 11× fewer on LongMemEval_S. Method and results: membase.com/research

Quickstart

[Hosted, local in process, or local over HTTP]

pip install unibaseio-membase              # hosted · Python 3.10+ · needs MEMBASE_API_KEYpip install 'unibaseio-membase[local]'     # local  · Python 3.12 or 3.13 · needs OPENAI_API_KEYpip install 'unibaseio-membase[media]'     # local + photos, videos and audionpm install @unibaseio/membase              # TypeScript · Node 18+npm install -g @unibaseio/membase-cli       # the `membase` command (hosted) without Python

from membase import Membase
m = Membase()                    # hosted# m = Membase(local=True)        # local, in ~/.membase
m.memories.add("We picked Postgres for the ledger service.")   # a factm.search("what database is the ledger on?")                    # {"items": [...], "partial": [...]}m.add("Design notes …", custom_id="design-1")                   # a documentm.profile_facts.list()                                         # who the user is (Your profile ticked)m.capabilities()                                               # what this key may do, and the rules

No call names a Memory: reads cover every Memory the key reaches, and writes land in its write target (a key that reaches several Memories and has none is refused with WriteTargetRequiredError). Ids are typed (mem_… a fact, doc_… a document, med_… a photo or video), lists are {items, has_more, partial}, and a removal without confirm=True raises ConfirmationRequiredError. Local memory answers the same routes with the same shapes.

Every method, its REST route and MCP tool, and the access it needs: SDK reference. Upgrading from 0.3: see the changelog. The Python package is unibaseio-membase (import membase); membase on PyPI is unrelated.

Command line

membase --local add "We picked Postgres for the ledger"membase --local search "what did we pick for the ledger?"membase --local memories listmembase --local ask "Which database is the ledger on?"membase --local import ~/Downloads/claude-export.json     # Claude, ChatGPT, markdown or JSON chatsmembase --local agent ingest trace.json --agent codermembase --local agent search "fix flaky deploy" --agent codermembase --local media add ~/Movies/trip.mp4                # shots, speech and speakersmembase --local media search "someone skiing"   # any language

Without --local (or --store DIR, both before the command) commands use the hosted API; agent, import and serve need unibaseio-membase[local], and media add and media ask (local only) need unibaseio-membase[media]: see Media memory. A local store with several Memories writes where MEMBASE_WRITE_TARGET names.

MCP

[MCP clients, hosted or local server, same tools]

claude mcp add membase -- membase --local mcp                                   # localclaude mcp add --transport http membase https://api.app.membase.com/mcp-http \  --header "Authorization: Bearer $MEMBASE_API_KEY"                             # hosted

Hosted without the header, the connection signs you in and is read-only. Setup for Claude, ChatGPT, Cursor, Codex and other clients: Connect your AI.

Gemini CLI, as an extension (it signs you in on first use; /mcp auth membase signs in again):

gemini extensions install https://github.com/unibaseio/membase

Cline: add the server to cline_mcp_settings.json; llms-install.md has the entry, with and without a developer key.

Docs

• SDK reference: methods, REST routes, MCP tools, access levels
• Memory operations: Memories, facts, documents, media, profile, ask
• Authentication: keys and what each one can reach
• Connect your AI: per-client MCP setup
• Agent framework integrations: LangChain, Microsoft Agent Framework, Dify, Hermes Agent, OpenClaw and Strands Agents
• Local engine: what it stores, membase serve, settings

License

MIT; see LICENSE. unibaseio-membase[local] also installs unibaseio-membase-core, the compiled engine, under its own Unibase license.