Membro

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.

Guarantees

• Facts are superseded, quarantined, or dismissed, never deleted by automation.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

Requirements

• Python 3.12+

• Optional: an Anthropic key (mining, summaries) and an OpenAI key (embeddings for semantic recall).

Quick start

git clone https://github.com/shawn-durrani/membro.git
cd membro
./start.sh

The 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-export

Imports are idempotent; re-running never duplicates messages.

MCP

claude mcp add -s user membro -e PYTHONPATH=<repo> -- <repo>/.venv/bin/python -m memory_service.mcp_server

Four 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.

Review queue

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.

Operations

Snapshots land in data/backups/ on a timer with rotation and an optional mirror folder (MEMORY_MIRROR_DIR). ops/install-supervisor.sh installs a launchd agent; restart with launchctl kickstart -k gui/$(id -u)/dev.membro.server. 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.

Docs

• 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.

Licence

MIT.