Skip to main content
Glama

cerebro-memory

A local, read-only index over your Markdown notes and your AI agents' memories, served over MCP, that tells you when they contradict each other.

Español

Why

You write a decision in your notes. Your coding agent saved an older version of it in its memory three months ago. A handoff doc in another repo says a third thing. Nobody notices until an agent confidently ships the old price, the old launch date or the feature you already killed.

cerebro indexes all of it on your machine (your Obsidian vault, Claude Code memories, repo docs and handoffs) and gives you, and any MCP client, one place to search it and one command to ask: where do my sources disagree?

  • Hybrid search that works in English and Spanish (and other languages): exact names via full-text search, paraphrases via multilingual embeddings.

  • Contradictions: pairs of fragments on a topic that state different prices, dates, versions or yes/no answers, with both sources and dates.

  • Local and read-only. Your files are never edited. Private notes never leave the index. Secrets are redacted before anything is stored.

Related MCP server: semdex

Quickstart (30 seconds)

You need uv.

uvx cerebro-memory demo

That copies a small fictional vault to a temp folder, indexes it and prints the contradictions it finds. The first run downloads the embedding model once; after that everything runs offline.

Then point it at your own notes:

uv tool install cerebro-memory        # installs the `cerebro` command
cerebro init --vault ~/Obsidian    # writes cerebro.toml; adds ~/.claude/projects if it exists
cerebro index                      # incremental; re-run any time
cerebro search "pricing decision"
cerebro contradictions "launch date" --across-sources

What it finds

Real output of uvx cerebro-memory demo (it prints the strongest disagreement per pair of notes; ask about one topic with cerebro contradictions "<topic>"):

$ uvx cerebro-memory demo
Demo vault copied to /tmp/cerebro-demo-x1y2z3/demo_vault
Indexed 26 notes from 3 sources (notes, agent-memory, repo-docs): 2 private (kept out of every result), 1 line(s) redacted, mode hybrid.

Your notes and your agents disagree on 3 things (older -> newer):

1. price: $4.99/month  ->  $5.99/month; price: $35/year  ->  $39/year; number: 7-day  ->  14 days
   older  2026-07-02  agent-memory:streak-web/memory/pricing.md
                      - Pro monthly: $4.99/month - Pro yearly: $35/year - 7-day free trial on both plans. Why: set when the pricing page was first built, from Lucía's brai…
   newer  2026-09-15  repo-docs:streak-web/docs/pricing-page.md
                      …Price | Free trial | |---|---|---| | Free | $0, up to 3 habits | none | | Pro monthly | $5.99/month | 14 days | | Pro yearly | $39/year | 14 days | …

2. version: 0.8.5  ->  0.9.3 / 0.9.0; date: December 1, 2026  ->  2027-01-12 / 2026-12-15
   older  2026-08-21  agent-memory:streak-app/memory/project-overview.md
                      …SQLite file is the source of truth; sync is optional and Pro-only. - Current version: 0.8.5 (beta channel). - Launch date: December 1, 2026. - Free …
   newer  2026-11-18  repo-docs:streak-app/docs/handoff-2026-11-18.md
                      - Current version: 0.9.3, in the beta channel since this morning. It fixes the midnight rollover in StreakClock.rollover() for UTC+9 and UTC-10. - La…

3. number: 7-day  ->  14 días
   older  2026-07-02  agent-memory:streak-web/memory/pricing.md
                      - Pro monthly: $4.99/month - Pro yearly: $35/year - 7-day free trial on both plans. Why: set when the pricing page was first built, from Lucía's brai…
   newer  2026-11-17  notes:Daily/2026-11-17.md
                      Martes tranquilo. Revisé las conversiones del beta: con la prueba de 14 días, 4 de 11 testers activaron Pro al terminarla. Me quedo con 14 días aunqu…

Run it on your own notes: cerebro init --vault ~/path/to/notes && cerebro index

The demo vault is in src/cerebro_mcp/demo_vault/; its README lists every fact planted in it.

Connect it to your agents

cerebro serve is an MCP server over stdio. Your client launches it; you never run it by hand. If you installed the tool, the command is cerebro serve; without installing, it is uvx cerebro-memory serve.

Claude Code

claude mcp add cerebro --scope user -- uvx cerebro-memory serve

Codex (~/.codex/config.toml)

[mcp_servers.cerebro]
command = "uvx"
args = ["cerebro-memory", "serve"]

Cursor (~/.cursor/mcp.json) and Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "cerebro": {
      "command": "uvx",
      "args": ["cerebro-memory", "serve"]
    }
  }
}

To use a config file other than the default, add "env": { "CEREBRO_CONFIG": "/path/to/cerebro.toml" }. More detail, file locations and Windows notes: docs/connect.md.

MCP tools

Tool

What it does

search

Hybrid search with filters (source, scope, type, since, until)

note

Full text, frontmatter and links of one note (by id, path or title)

links

Outgoing and incoming [[links]]

todos

Open checkboxes - [ ] in the notes that match a query

contradictions

Pairs of fragments on a topic that state different values

propose

Writes a Markdown proposal into cerebro's own folder; never touches your notes

Configuration

cerebro init writes a starter file. A full example:

[[source]]
name = "notes"
kind = "vault"                 # Obsidian/Markdown: frontmatter, [[links]], checkboxes
path = "~/Obsidian"
scope = "personal"
exclude = ["Templates/**"]

[[source]]
name = "claude-memory"
kind = "claude-code-memory"    # <path>/<project>/memory/*.md
path = "~/.claude/projects"

[[source]]
name = "handoffs"
kind = "glob"                  # any Markdown matched by the include globs
path = "~/code"
include = ["*/docs/**/*.md", "*/AGENTS.md", "*/CLAUDE.md"]
scope = "work"

[privacy]
private_folders = ["private", "people"]
private_frontmatter = "private"
redact = ["secrets", "cards"]

The config is looked up in this order: --config PATH, $CEREBRO_CONFIG, ./cerebro.toml, then your user config folder. Run cerebro doctor to see which file was used, how many files each source has and whether the model is cached.

Privacy model

  • Read-only. cerebro never edits, moves or deletes a source file. The only thing it writes besides its index is propose, into its own proposals folder.

  • Local. Parsing, indexing and embeddings run on your machine. The only network call is the one-time model download. Only the fragments a client asks for leave the process, and only to that client.

  • Private notes stay private. A note in a private folder (private/, people/ by default) or with private: true is indexed by path and hash only. Every query filters private notes out in SQL, so no tool can return them.

  • Redaction before storage. Lines with API keys, tokens, private keys or card numbers are redacted before the note is parsed. Packs: secrets, cards, and mx (Mexican IDs and bank numbers).

  • Hidden scopes. Sources in a hidden scope are indexed but only returned when that scope is asked for by name.

  • Content is data, not instructions. Every tool output carries a short notice saying so, and web clips (source: web) are marked untrusted.

Details and limits: docs/privacy.md.

How it works

  Markdown sources                     vault · agent memories · repo docs
        │
        ▼
  redact sensitive lines ─► parse frontmatter, [[links]], - [ ] checkboxes
        │
        ▼
  split into fragments by heading ─► skip unchanged files (content hash)
        │
        ├──────────────────────────────┐
        ▼                              ▼
  SQLite FTS5 (keywords, BM25)   sqlite-vec (multilingual embeddings, CPU)
        │                              │
        └──────────► RRF fusion ◄──────┘
                         │
                         ▼
        egress filter in SQL (no private notes, no hidden scopes)
                         │
                         ▼
          CLI  ·  MCP tools over stdio  ·  contradictions

Markdown is the truth; the index is disposable. Delete index.db and run cerebro index again: nothing is lost. More in docs/architecture.md and docs/contradictions.md.

Roadmap

v0.1 is the core: index, hybrid search, contradictions, MCP server, demo. Ideas for later, in no particular order and not promised:

  • Memory formats of more agents and editors.

  • A local model to judge contradiction candidates (today the client does).

  • A watch mode that re-indexes on file change instead of on a timer.

  • More redaction packs (contributions welcome).

  • contradictions limited to a date range or a project.

Contributing

Bug reports and pull requests are welcome. Read CONTRIBUTING.md first; security issues go through SECURITY.md, not public issues.

Credits

Built by César Méndez — surgeon and builder. cesarmendez.dev

Released under the MIT License.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      B
      maintenance
      Enables local AI agents to perform hybrid search (BM25 + vector) over privately indexed local files, including notes, PDFs, DOCX, source code, and OCR'd images. All indexing and retrieval stay on the machine with read-only, whitelist-scoped access.
      3
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables read-only semantic search over local folders of markdown, text, and optionally other document formats via a local vector store, without ever mutating the source files. It also watches each configured source for changes and auto-reindexes so search results stay fresh.
      123 PyPI
      MIT
    • F
      license
      Not graded
      quality
      A
      maintenance
      Exposes a local hybrid (vector + BM25) index of scattered notes, docs, PDFs, and chat logs to AI hosts like Claude Desktop, Cursor, and Cline, letting agents search and get cited answers. It publishes a set of tools, resources, and prompts so any MCP-compatible client can query that knowledge base with section-level citations.
      -
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables AI agents and MCP hosts to query a local, machine-resident index of scattered notes, project docs, PDFs, Office files, and chat exports using hybrid vector + BM25 retrieval with section-level citations. Exposes tools, resources, and prompts so hosts like Claude Desktop, Cursor, and Cline can search, ask, and explore linked knowledge through natural language.
      -