Midas

Local-first, eval-first memory for long-horizon AI agents — no LLM at ingest.

Midas is a small Python SDK (and an MCP server) that gives AI agents durable memory across long, multi-session work — coding agents, research agents, assistants — without sending every turn through an LLM to "extract" facts. It runs on your machine, costs nothing per message, and every recalled memory is traceable to its source.

• No LLM at ingest or query → $0 API spend, zero data egress, fast local ops (no per-turn network round-trip; ingest is embed-bound, ~tens of ms).
• Auditable provenance → recall returns the source turns, not LLM-rewritten facts.
• Stays current and bounded → belief revision, selective forgetting + tiers, dedup — all no-LLM.
• Embeddable + store-agnostic → a library, not a SaaS. Bring your own embedder/store.
• Eval-first → every claim has a reproducible benchmark (BENCHMARKS.md).

Status: early. The API may change. Built narrow and measured-first.

How it works (in plain English)

Your AI assistant forgets everything between sessions — every new chat starts from zero. Midas is a memory that lives next to your AI, on your computer. It does four simple things:

1. Notices what matters. As you work, Midas saves the durable stuff — a decision, a fact about you, a preference, a deadline — and ignores small talk. It judges "does this matter?" by reading the words (names, numbers, dates make a turn important) — without calling another AI.
2. Hands the right notes back. Before the AI answers, Midas finds the handful of past notes related to your question — by meaning, not exact keywords — and slips them into the prompt.
3. Keeps the notebook honest and tidy. When something changes ("actually, use Postgres now") it updates the old note instead of keeping both; it merges duplicates; and it forgets old, unimportant trivia so memory never bloats.
4. Stays yours. Everything is a local file — no cloud, no per-message AI bill — and every note links back to the exact moment it came from, so you can always check why the AI "knows" something.

The trick that makes it cheap, private, and local: Midas never sends your conversation to an AI to "process" it. It uses fast local math (embeddings — turning text into vectors and comparing them). The only AI involved is the one you're already talking to.

Why "no LLM at ingest" matters: other memory tools call an LLM to summarize every session — you pay in tokens forever, in latency, and by sending every turn to a provider. Midas trades that for cheap, local, auditable retrieval.

See it remember across sessions — session 1 stores decisions; a fresh session 2 recalls them by meaning:

_{Claude Code-style demo — the recalled lines (in green) are the real output Midas returned across two separate processes sharing one on-disk store.}

Install

You need Python 3.11+. Check with python --version (or python3 --version). If you don't have it: python.org/downloads, or winget install Python.Python.3.12 (Windows) · brew install python@3.12 (macOS) · your packa

... View full README on GitHub