Hi everyone, I'm Areg. I built iai-mcp because every memory tool I tried for AI coding assistants either forgot details I cared about or made me babysit what was worth remembering. I wanted neither. What it does: - Captures every turn of every session verbatim — byte-exact, not summarized — the moment the session ends. No "remember this" commands. - Consolidates during idle time into three tiers: episodic (write-once), semantic (induced summaries from clusters), procedural (sealed self-tuning knobs). - Surfaces a small relevant slice of your history at the start of each new session automatically. - Runs as a local daemon, encrypted at rest (AES-256-GCM), zero telemetry. The thing I care about most: every performance claim in the README ships with a runnable benchmark in `bench/`. Verbatim recall, latency, RAM, token cost, contradiction handling — all of it. Don't trust the numbers, rerun them on your machine. I've been running this daily for months. Validated end-to-end with Claude Code and OpenClaw. macOS only for now (Apple Silicon tested), Linux/Windows on the way. MIT licensed. Repo: github.com/CodeAbra/iai-mcp Happy to answer anything — architecture choices, the three-tier model, why the graph layer doesn't move the needle on session-shot retrieval benchmarks (long story), or implementation gotchas.

iai-mcp

The best open-source memory system for AI coding assistants.

The best open-source memory system for AI coding assistants.