🧠 Mnemo — local, token-free, graph-based project memory for Claude
Mnemo gives Claude durable project memory that is created and queried almost
entirely by local, free, open-source models — so it costs near-zero Claude
tokens to build or use. It turns a messy folder of documents (PDF, DOCX, PPTX,
XLSX, images, scans) into a knowledge graph, a compact Markdown digest, and
an interactive HTML mind map — and it is reusable across projects.
The idea: Claude tokens are expensive; local compute is free. So every heavy
step — document → text conversion, knowledge extraction, embeddings,
visualization — runs on your machine. Claude only ever issues a tiny tool call
and gets back compact metadata or a small, relevant subgraph — never whole
documents.
your documents ─► MarkItDown + Tesseract OCR + Ollama vision ─► Markdown (local)
│
Ollama qwen2.5:7b ── extract ──► entities · relations · facts
│
entity resolution (embeddings + fuzzy) ─► knowledge graph
│
┌──────────────────────────────┬───────────────────┴───────────────────┐
▼ ▼ ▼
graph.json memory.md mindmap.html
(source of truth) (compact digest Claude reads) (interactive Cytoscape graph)
│
▼
nomic-embed-text vectors ──► memory_query returns only a tiny relevant slice (not documents)
Contents
Why it saves tokens · What you get ·
Install · CLI ·
Claude / MCP tools ·
Reuse across projects · Configuration ·
Privacy · Resumable builds ·
Staying up to date · Apple-silicon optimization ·
How it works
Why it saves tokens
| Traditional "read the files into context" | Mnemo |
|---|
| Build | N/A (or paste docs → 100k+ tokens) | one local tool call → compact stats (~tens of tokens) |
| Use | re-read documents every time | memory_query → a few hundred tokens of the relevant subgraph |
| Where compute happens | Claude (paid tokens) | your machine (free, local) |
| Persistence | none (per-conversation) | global store, reusable across sessions & projects |
Measured on a real 62-file project (PDFs, DOCX, PPTX, spreadsheets, scanned
images): the converted source text is ≈ 440,000 tokens. Mnemo turns it into a
≈ 1,000-token memory.md digest (~450× smaller) plus a graph and mind map
— and the extraction itself runs on the local LLM, so the Claude-token cost to
build is ≈ 0. A typical memory_query answer is ~200 tokens. Loading full
project context this way costs ~1k tokens instead of ~440k.
What you get
- Knowledge graph (
graph.json) — entities, typed relationships, atomic facts,
each with provenance back to the source file.
- Compact digest (
memory.md) — overview + key entities/relationships/facts,
sized to a few hundred tokens; what Claude loads at session start.
- Interactive mind map (
mindmap.html) — a self-contained Cytoscape.js graph:
color-coded by type, searchable, click a node for its description, facts, and
sources. Works offline.
- Local retrieval — semantic search over the graph via
nomic-embed-text.
Requirements
macOS or Linux, Homebrew, and ~8 GB free disk for the models.
Everything else is installed for you. 16 GB RAM is comfortable for the default
qwen2.5:7b model (use qwen2.5:3b on smaller machines).
Install
git clone https://github.com/GRU-953/mnemo-mcp
cd mnemo-mcp
./scripts/install.sh # installs Ollama + Tesseract, pulls models, builds the venv
This installs/starts Ollama, pulls qwen2.5:7b (extraction),
nomic-embed-text (retrieval) and moondream (image captions), installs
Tesseract (OCR), and creates a Python 3.12 virtualenv with all dependencies.
Add --with-audio to also enable audio/video transcription (faster-whisper).
Use it from the command line
