Uses OpenAI's embeddings API to provide semantic recall over stored memories.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Membro What do you remember about my favorite books?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Membro is a local-first memory service for AI assistants. It ingests your conversations (including a claude.ai export), mines them for durable facts behind deterministic extraction walls, and serves the result to local models through a small HTTP API and an MCP server: ranked recall, verbatim history search, and a rebuilt profile summary with per-fact provenance. Everything lives in one SQLite file on your machine. Nothing listens beyond loopback unless you deliberately widen it to your own tailnet.
Facts are superseded, quarantined, or dismissed, never deleted by automation. The only hard delete is a button in the UI.
Every mined fact passes grounding, temporal, and source-trust checks before entering canon. Doubtful facts are held for your review.
Ingested messages are immutable. The summary is a cache rebuilt from the ledger, and each claim traces to its source facts.
With no API keys, ingest, search, and keyword recall still work; LLM paths raise clear errors instead of degrading silently.
Related MCP server: Memento
Python 3.12+
Optional: an Anthropic key (mining, summaries) and an OpenAI key (embeddings for semantic recall).
git clone https://github.com/shawn-durrani/membro.git
cd membro
./start.shThe server runs at http://127.0.0.1:8901; set MEMORY_PORT if that
port is taken. First run prints a recovery secret; use it once to enrol a
password, then log in with the password. Set MEMORY_AUTH_TOKEN in
.env to keep the secret stable.
To import your claude.ai export, unzip it first: the importer reads the
unzipped directory, not the zip. Stop the service before running it, or
point --data-dir at a throwaway copy. The only guard is a coarse one:
with no --data-dir, the importer quits if anything answers on
127.0.0.1:8901, and that port is hardcoded, so a service on another
MEMORY_PORT goes undetected. Pass --data-dir and there is no check at
all; it writes wherever you point it, running service or not.
unzip ~/Downloads/data-2026-08-01.zip -d ~/claude-export
.venv/bin/python scripts/import_claude_export.py --export-dir ~/claude-exportImports are idempotent; re-running never duplicates messages.
claude mcp add -s user membro -e PYTHONPATH=<repo> -- <repo>/.venv/bin/python -m memory_service.mcp_serverFour tools: recall_memory, search_history, save_memory,
memory_summary. MCP saves carry an mcp:* origin and are always
quarantined for review; agents cannot write into canon.
memory_service.mcp_admin_server is a separate opt-in read-only pair
authenticated with the bearer token.
Wall-flagged and untrusted facts wait in the web UI: approve, dismiss, quarantine, or supersede, all reversible. The miner scores importance 1 to 9; importance 10 never decays and only you can assign it. The Mathematics page shows the ledger as geometry and never shows fact content.
Snapshots land in data/backups/ on a timer with rotation and an
optional mirror folder (MEMORY_MIRROR_DIR). To go back to one, stop the
service and run scripts/restore_snapshot.py with the snapshot's path: it
keeps a copy of the live database first and replays every erasure made
since the snapshot, so a restore never brings back what you erased. The
fleet runbook (workbench runbooks/membro-restore.md) has the full
procedure.
ops/install-supervisor.sh installs a launchd agent; restart with
launchctl kickstart -k gui/$(id -u)/dev.membro.server. Before a
restart, GET /v1/busy says whether a job, a backup, a judge pass or a
re-embed is mid-flight; the fleet's deploy watcher waits on it. Everything
the service knows lives under data/; back that folder up and treat it
as sensitive.
Widening beyond loopback is tailnet-only and opt-in: read SECURITY.md first. Never expose the port to the open internet.
ARCHITECTURE.md: the settled decisions.
docs/MEMORY_DESIGN.md: the memory model.
docs/MEMORY_INTEGRITY.md: the extraction walls.
docs/API.md: the HTTP contract.
docs/TUNING.md: every knob and its default.
docs/TESTING.md: what the suite guarantees.
docs/REFERENCES.md: research lineage, including claims we checked and rejected.
MIT.
This server cannot be deployed
Persistent personal memory for AI assistants — save, search, and recall across every MCP client.
Persistent, portable memory for AI assistants — your private memory graph, from any MCP client.
Persistent, verbatim, searchable memory for AI assistants — one memory across every MCP client.
Persistent, inspectable memory for AI agents with lineage, correction, and a hosted MCP endpoint.
