memo

[memo — local memory for AI]

memo

Your coding agent starts every session with amnesia. memo fixes that — 100% on your own machine.

Persistent, searchable memory for Claude Code, Codex, Cursor, Cline, Devin, and OpenCode. No cloud, no API keys, no Ollama, no vector DB to run. And it spends fewer tokens, not more.

[PyPI] [Downloads] [License: MIT] [MCP] [MCP Toplist]

[mlx-memo MCP server on mcprush: security A, high access, no sign-in, works with 14 clients] [memo MCP server rated A on Glama: tool definitions rated A, 60 tools]

[Save a fact once — every later session recalls it automatically, all stored locally.]

Install

curl -fsSL https://raw.githubusercontent.com/jagoff/memo/v4.16.0/install.sh | bash

_{Prefer a package manager? uv tool install mlx-memo · pipx install mlx-memo · brew tap jagoff/memo && brew install mlx-memo}

Then:

memo doctor                                   # self-checkmemo save 'we use Postgres, not Mongo'        # save a decisionmemo search 'what database did we pick?'      # search by meaning

That's it. Your agents pick it up over MCP automatically — the installer wires every client it finds.

curl -fsSL https://raw.githubusercontent.com/jagoff/memo/v4.16.0/install.sh | bashmemo sync bootstrap [email protected]:yourname/memo-sync.git

curl -fsSL https://raw.githubusercontent.com/jagoff/memo/v4.16.0/install.sh | bashmemo doctor --strict-runtime

On Linux or just want to look around first?

docker run --rm ghcr.io/jagoff/memo:latest memo doctor

Why this saves you money

Most memory servers add context. memo is built to remove it.

The default MCP surface is 43 tools, not 165 — 74% fewer tools, and about 68% less schema context: 43 tools / ~9.7k schema tokens versus 165 tools / ~30.6k tokens on the full surface — overhead paid every session, in every client.

Ambient recall injects one relevant memory before the model answers. The bundled Claude Code hook caps that injection at ~160 tokens. memo roi reports the real grounding and re-ask counts — the estimated-savings figure it used to print was removed in 4.14.0, because multiplying those counts by hardcoded constants was a savings claim memo could not support. For measured savings, memo tokens reads the provider's own usage counters through the context-compression proxy.

memo roi       # value from grounded recalls and avoided re-asksmemo tokens    # usage-savings ledger

Three things nothing else does

🕰️ Time-machine — query your knowledge as it was

memo as-of ask "what was the deploy strategy?" --date 2026-02-01memo diff --from 2026-01-01 --to 2026-03-01

Full historical reconstruction by reverse-replaying history.db. Useful when you need to know why past-you made a call, not just what past-you decided.

⚡ Contradiction radar — memory that notices when you change your mind

memo contradict scan      # find conflicting facts corpus-widememo contradict triage    # resolve: fuse / newer-wins / dismiss

Change a decision and memo flags the now-stale version, so the agent stops reintroducing what you already threw out.

🔮 Dream — it optimizes itself while you sleep

memo dream run

A 7-phase nightly pipeline: inventory → mine signals → resolve conflicts → prune stale → synthesize cross-cluster insights → optimize → pre-warm the top-100 query embeddings so tomorrow's recall stays under 200 ms. Every run writes a receipt you can audit. Zero intervention.

How it works

Hybrid retrieval. A vector leg (MLX on Apple Silicon, sentence-transformers on CPU) and a BM25 leg (FTS5, diacritic-folding for Spanish) run in parallel, fuse via Reciprocal Rank Fusion, then go through an optional MLX cross-encoder rerank.

[vector + keyword search in parallel, fused, reranked, top memory injected]

Markdown is the source of truth. Every memory is a plain .md file you can read, grep, and version-control. SQLite is a derived index that rebuilds from the files at any time — hand-edit in Obsidian and your edit wins on the next memo reindex. Nothing is locked in a database you can't open.

Prompts and memories stay on your machine. Embedder, reranker, and LLM all run in-process. No telemetry. Memory travels only if you point memo sync at a git remote you own. Normal startup is fully offline; remote update checks and auto-update require an explicit opt-in. → Privacy and network policy

Also in the box: cross-agent memo resume (reopen any session from any agent), cross-Mac git sync, a knowledge graph with optional codegraph symbol edges, encrypted secret storage, OCR/audio ingestion, evidence packs, outcome learning, signed federation, and a local chat UI over your memory (memo chat serve). → Full feature reference

How it compares

Verified July 2026 against each project's own docs. Corrections welcome — open an issue and I'll fix the table.

_{✅ first-class · ⚠️ partial, config-gated, or add-on · ❌ absent}

Closest comparators are basic-memory (local-first + Obsidian + MCP — same thesis) and cipher (memory for coding agents).

Requirements

Python ≥ 3.13 (the installer handles this via uv if you don't have it). First install pulls ~8 GB of models, 5–15 min. Optional: an Obsidian vault — without one, memo uses ~/Documents/memo/.

Docs

All 145 top-level CLI commands

Contributing

git clone https://github.com/jagoff/memo && cd memouv pip install -e '.[dev]'

Issues and PRs welcome — see CONTRIBUTING.md. If memo is useful to you, a ⭐ genuinely helps other people find it.

MIT licensed. Built on Apple MLX, sqlite-vec, and codegraph.