Origin

AI work memory for Claude Code. Origin carries sessions, decisions, lessons, and project context forward, then turns them into searchable memory and wiki pages.

30-Second Setup

0s   /plugin marketplace add 7xuanlu/origin
     /plugin install origin@7xuanlu
5s   restart Claude Code
10s  /init   auto-installs daemon if missing, configures local memory,
            verifies daemon + MCP + round-trip, prints "Origin ready"
30s  /brief  (or /capture <something to remember>)

/init is self-healing — if the daemon isn't running and the origin CLI isn't on PATH, it runs the install one-liner for you. No copy/paste, no restart loop. The SessionStart hook only nudges you toward /init if the daemon ever stops.

Install

/plugin marketplace add 7xuanlu/origin
/plugin install origin@7xuanlu

7xuanlu is the GitHub repo owner. If you fork Origin, use your own handle.

The marketplace is defined in ../../.claude-plugin/marketplace.json (at the repo root). The plugin metadata is defined in plugin.json. MCP configuration is in ../.mcp.json (this plugin's .mcp.json), which delegates to ../bin/origin-mcp-runner.sh.

The runner picks the MCP server binary in three paths, in order:

1. Filesystem override — if plugin/bin/origin-mcp.local exists (typically a symlink to a locally-built binary, gitignored), the runner exec's it. Most reliable: survives plugin reloads that don't re-read env.
2. Env var override — ORIGIN_MCP_DEV_BIN=/abs/path/to/origin-mcp. Convenient if you already export it; requires Claude Code to inherit the var at startup.
3. Default — npx -y origin-mcp@^X.Y.Z. What end users get after installing the plugin.

To set up the filesystem override during dev:

cargo build -p origin-mcp --release
ln -s $(pwd)/target/release/origin-mcp plugin/bin/origin-mcp.local

Reload the plugin (/reload-plugins) and the wrapper picks the local binary on the next MCP spawn.

Daily Commands

/init       set up + verify Origin works (run once, or to diagnose)
/help       one-screen reference
/brief      load identity + topic context (start of session)
/capture    save one durable memory in flow
/recall     search local memory
/distill    synthesize pages from clusters (scoped to current repo)
/read       preview a distilled page inline
/review     audit pending memories
/forget     delete a memory by ID
/handoff    end-of-session debrief
/debrief    alias for /handoff (brief/debrief symmetry)

A SessionStart hook (hooks/check-daemon.sh) probes the local daemon at 127.0.0.1:7878. If down, it prints a single line: daemon not running. Run /origin:init to set up. The skill owns the install logic — the hook is just a nudge. Hook never blocks the session.

Where your data lives

~/.origin/pages/               wiki pages distilled from memories (md)
~/.origin/sessions/            session logs by date (md)
~/.origin/sessions/_status/    current per-project goals + last-handoff timestamp
~/.origin/db/                  symlink to the libSQL store
~/.origin/bin/                 installed binaries

Browse with open ~/.origin/ (Finder), code ~/.origin/ (VS Code), or symlink ~/.origin/pages/ into an Obsidian vault for the graph view. No Tauri app required.

Local memory and agent-side model phases

By default /init configures local memory: no model download, no API key, no prompts. The daemon stores, embeds, dedupes, and serves hybrid search. Model-backed work like classification, entity extraction, page synthesis, and reranking stays opt-in.

The skills compensate. Where the daemon would normally call a model, the skill asks Claude itself to do the equivalent step and posts the result back via HTTP:

Phase	Model-equipped daemon	Local memory + skill
Pick memory_type	daemon classifier	/capture picks one of 6 types from content
Extract entities/relations	daemon extract.rs	/capture POSTs to /api/memory/entities and /api/memory/relations
Synthesize a page	daemon distill cycles + /distill	/distill reads the cluster, writes the page, POSTs to /api/pages
Expand a query / rerank hits	daemon expansion + rerank	/recall rewrites the query before search and reorders hits after

The daemon stays the single writer and storage owner. Claude does the thinking. This is why local memory works without an API key or on-device model — and why the same skills also work when you later add one. The skills detect the daemon's capabilities and fall through to the agent path when needed.

Skill Files

The actual skill instructions live in ../skills:

• init: end-to-end setup verifier (daemon + MCP + round-trip)
• help: one-screen quick reference
• brief: load session context
• capture: save one durable memory
• recall: targeted lookup
• distill: refresh wiki pages
• read: preview a distilled page inline
• review: audit pending memories
• forget: delete a memory by ID
• handoff: capture end-of-session decisions, lessons, gotchas, and open threads
• debrief: alias for handoff (brief/debrief symmetry)

Links

• useorigin.app — project home
• useorigin.app/learn/claude-code-memory — Claude Code memory concept article
• useorigin.app/docs/daily-workflow — the brief/capture/recall/handoff loop
• useorigin.app/docs/get-started — install + verify

License

Apache-2.0.