Notes your AI agent
can actually use.
npx mnotes connect claude — then your agent reads, edits, and creates notes. Works with Claude Code and Cursor.
Works with Claude Code · Cursor · Any MCP client
Everything you need.
Nothing you don't.
A markdown editor, semantic search engine, and AI integration layer — in one app.
Built-in MCP Server
Claude, Cursor, and ChatGPT read and write your notes through a standard Model Context Protocol endpoint.
Semantic Search
Find notes by meaning, not just keywords. pgvector embeddings surface the most relevant context automatically.
Beautiful Editor
CodeMirror 6 with live preview, syntax highlighting, vim keybindings, and a dark UI built for long sessions.
Smart Organization
Nested folders with drag-and-drop. Structure scales from 10 notes to 10,000 without friction.
Private & Secure
Your data stays yours. OAuth with Google and GitHub, email/password auth, and encrypted storage.
Knowledge Graph
Visualize how your notes connect. Wikilinks build a force-directed graph — nodes sized by importance, edges by relationship.
Top-tier recall on LongMemEval-S.
ICLR 2025 benchmark for long-context memory retrieval. m-notes' hybrid search (Postgres FTS + pgvector) lands within a point of the state of the art — and well ahead of established frameworks.
| System | R@5 | R@10 | NDCG@10 | MRR |
|---|---|---|---|---|
| m-notes (hybrid)this app | 90.0 | 96.0 | 85.7 | 82.3 |
| agentmemory | 95.2 | 98.6 | 87.9 | 88.2 |
| Letta | 83.2 | — | — | — |
| Mem0 | 68.5 | — | — | — |
Metric: recall_any@K — fraction of questions where at least one gold-evidence session appears in the top-K retrieved results. m-notes measured on a 50-instance smoke run of LongMemEval-S (Xiaowu et al., ICLR 2025) using bge-m3 (1024-d, local) + Postgres FTS; competitor numbers from each project's published report. Full 500-instance result lands when complete.
Your AI reads your notes.
Then writes new ones.
Via MCP, your AI assistant searches notes by meaning, recalls past decisions, and creates new knowledge — all through standard tool calls.
Not just an MCP tool.
A full note-taking app.
A beautiful markdown editor you actually want to use — with folders, search, and a dark UI built for focused work.
Markdown-native editor
CodeMirror 6 with live preview, syntax highlighting, and vim keybindings.
Nested folders
Organize with drag-and-drop. Scales from 10 to 10,000 notes.
Hybrid search
Full-text + semantic search. Find by keywords or meaning.
Dark-first UI
Obsidian-inspired interface built for long coding sessions.
See how your ideas
connect.
Link notes with [[wikilinks]] and watch a force-directed graph emerge. Nodes grow with importance — the more connections, the bigger the node.
Type [[note title]] to create bidirectional links between notes.
Nodes scale with connection count — important ideas stand out.
Hover to highlight clusters. Click any node to open the note.
Three steps to agent memory
Connect once. Store knowledge as you work. Recall it in every future session.
Connect
One command wires up MCP. Your AI agent connects to m-notes in seconds.
$ mnotes connect claude-codeMCP config written. Agent connected.
Store
Your agent stores decisions, context, and learnings as it works. Knowledge persists across sessions.
knowledge_store({content: "Use pgvector for embeddings",tags: ["architecture"]})
Recall
In a new session, your agent recalls relevant knowledge by meaning. Context is never lost.
semantic_search({query: "vector storage decision"}) // score: 0.94
Built for developers.
By developers.
Modern stack, type-safe APIs, built for integration.