📣 Now open source

OpenLTM was born as a private memory layer for Claude Code. Today it's fully open source under MIT.

Same engine — automatic capture, semantic recall, importance-weighted decay, a queryable memory graph — now yours to read, fork, and extend. What started as one developer's "stop re-explaining my codebase" plugin is now an open foundation for agent memory across Claude Code, OpenCode, and Pi.

• 🔓 MIT licensed — no cloud, no account, no telemetry. Your memory lives in a local SQLite DB you own.
• 🧩 Provider-agnostic — one core (@rohirik/openltm-core), thin adapters per host.
• 🛠 Hackable — hooks, skills, janitor providers, and the graph visualizer are all in the open.

Migrating from an earlier install? The marketplace is now RohiRIK/OpenLtm and the plugin is openltm. Your existing memory database carries over.

The philosophy

Four ideas. No exceptions.

• Memory should be automatic. Hooks do the work. The session end hook extracts patterns, the session start hook injects them back. You shouldn't have to remember to remember.
• Decay is a feature, not a bug. A gotcha from six months ago that you never revisited probably no longer applies. Set importance: 5 to make something permanent — everything else ages out naturally.
• Semantic over keyword. FTS5 full-text search runs first; if it returns nothing, vector embeddings kick in. You search by meaning, not exact words — "how we handle async errors" finds the right memory even if you never wrote those exact words.
• Zero config, zero lock-in. Install once, works everywhere. Every setting has a sane default. The DB lives outside the plugin directory so it survives every update. No cloud, no telemetry, no account.

What you get

Install

Marketplace (recommended)

claude plugin marketplace add https://github.com/RohiRIK/OpenLtm
claude plugin install openltm

Restart Claude Code. That's it. Four Claude Code hooks + one git post-commit hook auto-wire, four commands load, five skills activate, and your openltm.db migrates or creates itself.

bunx (no clone)

bunx @rohirik/openltm-core          # auto-detect Claude Code, OpenCode
bunx @rohirik/openltm-core --pi     # experimental Pi adapter
bunx @rohirik/openltm-core --dry-run --claude  # preview without writing

Dev / git clone

git clone https://github.com/RohiRIK/OpenLtm ~/Projects/OpenLtm
cd ~/Projects/OpenLtm && bash install.sh

Quick Start

Start a new session. Context is injected at the top automatically.

Then try:

/openltm:memory recall auth       — what do we know about auth in this project?
/openltm:memory learn <insight>   — save something worth keeping
/openltm:health                   — memory health + decay summary
/openltm:project init             — set a goal for the current project

That's it. The rest is hooks doing the work.

The shape of memory

Claude Code
   │
   ├── 4 Commands  ──┐
   ├── 5 Skills    ──┼──▶  openltm MCP server ──▶  openltm.db
   └── 5 Hooks     ──┘                        (memories, tags,
                                                context_items,
                                                memory_relations,
                                                memories_fts)

Full deep-dive — schema, hook architecture, decay formula, ADRs — in How It Works and Architecture.

SQLite Extensions: The plugin loads sqlite-vec (vec0 / KNN vector search) and Honker (async embedding queue, leader-elected cron, pub-sub) when a system extension-enabled libsqlite3 is available. Both degrade gracefully — missing binary or library leaves the capability off and falls back to JS-cosine, file-watch polling, or in-process cron. Controlled with LTM_DISABLE_VEC, LTM_DISABLE_HONKER, LTM_SQLITE_LIB, and LTM_HONKER_EXT env vars.

Verify

/openltm:health                    # plugin health + hooks + decay
/openltm:memory recall test        # returns results (or "no results" on fresh install)

Start a new session — you should see context injected at the top. If not, run /openltm:health to diagnose.

Go deeper

Full documentation index: docs/.

Contributing

Open an issue first to discuss the change, then send a PR. The full workflow — setup, tests, the version-bump rule, and release steps — is in CONTRIBUTING.md.

Report a Bug

License

MIT — RohiRIK