levh

What is LEVH?

Your AI tools are stateless — every session starts from zero. LEVH gives them a persistent, searchable memory that lives on your machine. Plug it into Claude Desktop, Cursor, Claude Code, VS Code, or any MCP client, and it remembers your decisions, projects and people across sessions.

Most memory tools store everything and retrieve everything. LEVH forgets on purpose: every memory has its own decay curve, unused ones fade, and the ones you actually rely on are reinforced automatically. Signal rises to the top without manual curation.

Everything runs locally on SQLite. No accounts, no cloud, no external services.

Related MCP server: trw-mcp

Quick start

pip install levh
levh setup --client claude --profile work
levh capture "Atlas uses PostgreSQL in production."
levh serve

Dashboard and API come up on http://localhost:8000. Open Settings for copy-paste MCP configs, or run levh mcp config cursor for any supported client.

Optional, once you are set up:

levh hook install --client claude-code   # every new session starts with your memory
levh hook install                        # capture every git commit message
levh context -o CLAUDE.md                # compile memory into a context file

How it works

Every memory has its own half-life. Recalling it resets the clock and grows it, negative feedback shrinks it, and pinning skips decay entirely.

retention(t) = 0.5 ^ (hours_since_last_recall / stability_hours)

Memories are ranked by an explainable multi-factor score — semantic similarity, decay, importance and access frequency, each weight configurable. See Architecture.

What you get

• Adaptive decay — per-memory half-life, reinforced by recall and weakened by feedback.

• Ask your memory — natural-language answers that cite the memories they came from. Offline by default.

• People, projects & timeline — entities extracted from your memories automatically. No manual tagging.

• Daily briefing — what's on today, what you recently committed to, and what you're about to forget.

• Decisions & conflicts — decisions pulled from your memories, with a review signal when two disagree. A signal, never a verdict.

• Knowledge Wiki — a source-linked Markdown wiki built from your memories, browsable in the dashboard. It surfaces the conflict candidates LEVH already flagged, each linked to both snapshots — a candidate, never a verdict.

• Admission gate — every memory screened before storage; duplicates flagged and secrets redacted. An ambiguous candidate is held for you, never dropped.

• Trust & provenance — an explainable reliability score per memory. Separate from ranking; not a truth claim.

• Mistake guard — a corrected mistake becomes a pinned rule that never decays.

• Encrypted backup & restore — one portable snapshot, optionally encrypted.

• Connectors — Calendar, Email, transcripts, Notion, Obsidian, GitHub, Jira, Linear, Slack, git history and local files.

• Signed federation — explicit, pull-first memory exchange between two LEVH instances.

• 78 MCP tools, a REST API, a WebSocket feed, and a Next.js dashboard served by the API itself — one process, one port.

Documentation

→ Installing from source · Configuration

Security

LEVH is a local, single-user tool — no accounts, no multi-tenancy.

• Tokenless means loopback-only. Without LEVH_TOKEN, remote peers are rejected.

• LEVH_TOKEN gates /api/*, the WebSocket and the MCP SSE transport. Set it before widening any bind. Generate one with python -c "import secrets; print(secrets.token_urlsafe(32))"; it is a value you choose, not one the server hands out.

  • If the dashboard asks you for a token, paste the same value you started the server with — whatever LEVH_TOKEN held in the shell, .env or Docker environment that ran levh serve. The server never prints it, so it is not shown anywhere; if you never set one, pick a value with the command above and restart with LEVH_TOKEN=<value> levh serve. Once set, every call needs it, including from the same machine — the token is what the dashboard is gated on, not the network distance.

• Nothing leaves the machine without an explicit opt-in. Ask, summaries and transcript ingest stay offline until you set ANSWER_MODE=llm or SUMMARY_MODE=llm.

• Secrets are redacted before storage, on every write path.

The MCP size profiles narrow tool discovery; read_only is enforced server-side at call time. The size profiles are not a security boundary.

Found a vulnerability? See SECURITY.md.

Contributing

Issues and pull requests are welcome — see CONTRIBUTING.md and Discussions.

License

GNU Affero General Public License v3.0 or later (AGPL-3.0-or-later).