omem

Memory for AI agents that lives in plain markdown.

You read and write it in Obsidian. Agents read and write it over MCP. Git keeps it in sync. No LLM, no cloud, no lock-in — just markdown, an index, and a server.

Install

# one-shot: install, pick a vault, index, register with your MCP client
npx -y @kipachu/omem setup

# or manual
npm i -g @kipachu/omem
omem init ~/my-vault   # template vault (or point at an existing one)
omem setup             # wire it up
omem agents            # register with Claude Code, Cursor, Windsurf, Gemini CLI, …

No vault yet? template/ is a ready-to-use starting structure: per-domain islands/, inbox/ for triage, archive/ for superseded notes, and CONVENTIONS.md that teaches agents the write rules.

npm version: published latest is 0.7.2 — the core feature set below is live on the registry.

How it works

              ┌──────────────┐        ┌──────────────┐
              │  Your vault  │ ◀────▶ │  Obsidian    │  (you)
              │  (markdown)  │        └──────────────┘
              └──────┬───────┘
                     │ chokidar poll
                     ▼
              ┌──────────────┐
              │  indexer     │── FTS5 + vector cosine (ONNX) + 1-hop wikilink graph
              │  (SQLite)    │   →  fused via RRF + recency + pinned/kind boost
              └──────┬───────┘
                     │
                     ▼
              ┌──────────────┐        ┌──────────────┐
              │  omem serve  │ ─MCP─▶ │  AI agent    │  (Claude Code, Cursor, …)
              │  (stdio/HTTP)│ ◀────  └──────────────┘
              └──────┬───────┘
                     │ git pull / push + commit
                     ▼
              ┌──────────────┐
              │  GitHub repo │  (your vault remote)
              └──────────────┘

The vault is the single source of truth. The SQLite index at .omem/index.db is fully derived — delete it anytime, omem rebuild regenerates it. Retrieval is hybrid: FTS5/BM25 keyword + brute-force vector cosine + 1-hop wikilink graph expansion, fused with Reciprocal Rank Fusion.

No LLM and no external services in the pipeline. Embeddings run in-process via transformers.js (ONNX); the model downloads once (~30 MB) and works offline after.

Why omem

Most agent-memory tools are vector-DB-first. omem is vault-first:

• Your memory is plain markdown. Read, edit, search, and back it up with stock tools (Obsidian, vim, git, grep). No proprietary format, no export dance.
• You and your agents share one store. When you write a note in Obsidian, the agent sees it on the next pull. When the agent writes a note, you see it in Obsidian on the next push. No "agent memory" vs "human memory" split.
• No LLM in the pipeline. Embeddings run in-process (ONNX, ~30 MB, offline). Extraction, summarization, routing — those are an LLM call away if you want them, but omem never calls one for you.
• Git is the audit log + sync. Every write is a commit. Every pull is a rebase. Conflicts resolve local-wins; the other side stays recoverable in git log.
• Obsidian is the UI. Graph view, backlinks, daily notes, plugins — your existing Obsidian workflow, unchanged.

What an agent gets

An agent that connects to omem gets a memory it can orient, recall, read, write, and refine — all over MCP, all against your markdown vault.

• Orient. Land on a fresh session and learn the vault in one call: how many notes, what folders, what tags, what's recent. No guessing.
• Recall. Hand it a task or question; get back ranked results grouped by kind — decisions, gotchas, conventions float to the top. Pinned facts rank first.
• Search & read. Hybrid keyword + vector + graph search over every note. Full notes with backlinks. Browse by folder or tag without a query.
• Write & refine. Agents write plain markdown with frontmatter. Before writing, they see near-duplicate candidates so they append instead of creating dupes. Notes can be stamped with a kind (decision, gotcha, convention, …) and pinned for canonical facts. Superseded notes are archived, never deleted.
• Observe. Per-client watermarks mean an agent can ask "what changed since I last looked" and get a focused answer, not the whole vault.

The server ships instructions on the MCP initialize handshake — the nudge to recall before acting travels into the agent's system prompt automatically, on every session. (src/mcp/shared.ts, ≤400 chars, test-guarded.)

Quick demo

Once omem serve is running and your MCP client is connected, an agent lands on a task and the server instructions tell it to recall first. Real output from a fresh template/ vault:

Orient — one-call vault snapshot:

{
  "notes": 7, "chunks": 16, "lastModified": "2026-07-06T23:11:59Z",
  "topFolders": [{"folder": "islands", "count": 3}],
  "pinned": 0, "archived": 1,
  "recent": [{"path": "CONVENTIONS.md", "title": "CONVENTIONS",
              "link": "obsidian://open?vault=…&file=CONVENTIONS"}]
}

Recall — context-in, ranked, with clickable obsidian:// deep-links:

{
  "query": "what conventions should I follow when writing notes?",
  "grouped": {"decision": [], "gotcha": [], "convention": [],
              "fact": [], "meeting": [], "log": []},
  "related": [
    {"notePath": "CONVENTIONS.md", "title": "CONVENTIONS",
     "heading": "Conventions", "score": 0.0164, "matchType": "keyword",
     "link": "obsidian://open?vault=…&file=CONVENTIONS"},
    …
  ],
  "totalScanned": 9
}

Write — creates the note and returns dedup candidates so the agent appends instead of duplicating:

{
  "path": "islands/example-project/2026-07-06-demo-decision.md",
  "mode": "create",
  "link": "obsidian://open?vault=…&file=islands%2Fexample-project%2F2026-07-06-demo-decision",
  "similarExisting": [
    {"path": "archive/README.md", "score": 0.857},
    {"path": "inbox/README.md", "score": 0.844},
    {"path": "CONVENTIONS.md", "score": 0.835}
  ]
}

Run modes

Local stdio (single agent, same machine): omem serve --vault ~/my-vault HTTP (remote agents or Railway deploy): omem serve --port 8080 --vault ~/vault Set OMEM_HTTP_TOKEN for HTTP auth — without it the endpoint is open. Never expose an unauthenticated port publicly.

Conventions

Every note needs YAML frontmatter. The full schema lives in template/CONVENTIONS.md. The two rules that matter:

1. Search before writing. Agents recall before acting and append to existing notes instead of duplicating.
2. Never delete. Superseded notes are archived, not removed. History survives in git.

Deploy (Railway / Docker)

The repo ships a Dockerfile + start.sh that run a 24/7 memory server: the vault is cloned at boot, served over HTTP, git-synced both ways.

1. Create a Railway service from this repo (Dockerfile auto-detected).
2. Mount a volume at /vault — persists the clone, index, and ONNX model.
3. Set env vars:
   • VAULT_REPO — e.g. youruser/your-vault
   • GITHUB_TOKEN — fine-grained PAT, read/write contents on that repo only
   • OMEM_HTTP_TOKEN — openssl rand -hex 32
4. Generate a public domain. On each client:

   claude mcp add --transport http omem https://<app>.up.railway.app/mcp \
     --header "Authorization: Bearer $OMEM_HTTP_TOKEN"

Related

• template/ — ready-to-use starting vault.
• src/mcp/shared.ts — the instructions string agents see.
• Model Context Protocol — the transport.
• Obsidian — the human UI.

License

MIT © Kipachu.

Contributing

See AGENTS.md for repo conventions. Small PRs, conventional commits, one feature per PR.