membase-ai

One memory. Every AI.
Drop-in memory infrastructure for AI agents and apps — the Membase SDK.

Website · Docs · Web app · Discord · X

Membase gives agents and apps context that persists, built for production. This repository is its SDK for developers: the Python client, the membase command line, an MCP server and the TypeScript client, for memory hosted in your Membase account or kept on your machine. Unibase Memory, the product for people (the Chrome extension, the web app and the desktop app), is powered by Membase.

How it works

1. Install the SDK: pip install membase-ai or npm install membase-ai.
2. Integrate: connect any MCP-compatible agent with one command (see MCP), or call the client from your code.
3. Manage: search, edit and organise what the memory holds from the web app, or with membase on the command line.
4. Retrieve: your agent pulls the relevant context for every request (search), or asks the memory for an answer (ask).

Benchmarks

Accuracy and context efficiency of the engine, measured end to end with membase-bench:

Install

The methods and response shapes are the same either way; code moves between hosted and local memory by changing the constructor. TypeScript: npm install membase-ai (typescript/). Install membase-ai, not membase: that is an unrelated package with the same import name.

from membase import Membase

m = Membase()                    # hosted: MEMBASE_API_KEY (Connect › Developer keys)
# m = Membase(local=True)        # or local: ~/.membase

m.memories.add("We picked Postgres for the ledger service.", container="Engineering")
m.search("what database is the ledger on?")
m.add("Design notes …", container="Engineering", custom_id="design-1")   # a document
m.profile()
m.ask("Which database did we choose for the ledger?")   # hosted, this needs an agent-endpoint key

Every method is one operation of the Membase agent protocol (list_containers, search_memories, get_profile, list_documents, get_document, memory_rules, add_memory, add_document, delete_document, forget_memory, ask_agent). Hosted, the service enforces each key's reach and access level. Local, the same operations run on the membase-core engine: a memory becomes dated episodes, a document becomes a topic tree, and search is the engine's multi-round retrieval. Removing a document or forgetting a memory needs confirm=True in both.

Command line

membase --local add "We picked Postgres for the ledger" --container Engineering
membase --local search "what did we pick for the ledger?"
membase --local ask "Which database is the ledger on?"
membase --local import ~/Downloads/claude-export.json     # Claude / ChatGPT / markdown / JSON chats
membase --local import ~/chats/ --dry-run                 # a directory, scanned; --dry-run only reports
membase --local documents add notes.md --container Engineering
membase --local profile
membase --local agent ingest trace.json --agent coder     # agent memory: traces -> cases and skills
membase --local agent search "fix flaky deploy" --agent coder

--local and --store DIR go before the command. Without them the commands use the hosted API (MEMBASE_API_KEY), except agent and serve, which are local only; import needs membase-ai[local] either way to read the exports, and hosted it stores each chat as a document. MEMBASE_LOCAL=1 (or a directory) makes local the default when MEMBASE_API_KEY is not set.

MCP

claude mcp add membase -- membase --local mcp       # local memory
claude mcp add --transport http membase https://api.app.membase.io/mcp-http \
  --header "Authorization: Bearer $MEMBASE_API_KEY"  # hosted, at the key's access level

Hosted without the header, the client signs in and the connection is read-only. The local server offers the hosted endpoint's agent-protocol tools: list_containers, search_memories, get_profile, list_documents, memory_rules, add_memory, add_document, delete_document, forget_memory, ask_agent.

Local memory over HTTP

membase --local serve            # http://127.0.0.1:8787/v1

membase serve answers the agent-protocol routes of the hosted API (containers, search, profile, rules, documents, memories, ask) from the local engine, so the TypeScript client, or anything else that speaks the API, can use local memory by pointing its base URL at it. With MEMBASE_LOCAL_TOKEN set, every request must carry that token as its bearer.

Local settings

Local stores: the default container is ~/.membase/memory.db, others are ~/.membase/containers/<id>/, and standing (static) profile facts are in ~/.membase/profile/. The engine reads its settings from the environment, and from a .env file in the current directory or a parent without overriding what is already set:

Where things live

membase-sdk (PyPI and npm) is the earlier name of this package and is kept as an alias.

License

MIT, for this repository: the client, command line, MCP server and TypeScript package. membase-ai[local] also installs the compiled membase-core engine, which is under Unibase's proprietary license.