iai-personal-memory-engine

Give your AI a memory that never forgets

Open Source

Developer Tools

Artificial Intelligence

GitHub

Visit WebsiteSee on Product HuntGithub ⧉

Hey Product Hunt 👋 Here's the thing nobody tells you about AI coding agents: forgetting is expensive. When your agent doesn't remember something, it goes hunting — greps the repo, opens five files, reburns thousands of tokens to rediscover a fact you told it last week. On my own store a search round-trip averages ~2,850 tokens. Serving that same fact from memory costs ~350 — about 88% cheaper, every single time. Over three recent weeks of normal work, that was 282 memory packs served, ~99,000 tokens of context injected, and ~707,000 tokens of agent searching never paid for — by the engine's own deliberately conservative formula. The dashboard keeps your own count, live, so you don't have to take my number for it. What it actually is: iai-pme captures every turn verbatim and write-once. Most memory tools summarise — and summaries lose the one sentence you actually needed. When a fact changes here, the old version is archived, not overwritten, so both stay retrievable. That scores Rescue@10 = 1.000 on our post-contradiction benchmark, and R@5 = 0.962 on LongMemEval-S. All of it local. AES-256-GCM at rest, embeddings computed on your machine, no account, no API key, no telemetry. The only thing that leaves your laptop is the model call your CLI already makes. Fifteen MCP tools, including recall that returns contradictions alongside matches — so a stale fact can't quietly pass itself off as current. Every number above ships with the harness that produced it. Clone it, run python -m bench.longmemeval_blind, and check me. pip install iai-pme Works with Claude Code and Codex out of the box (ambient capture), plus Gemini CLI, Cursor and anything else that speaks MCP. macOS and Linux; Windows is in beta. MIT. Happy to answer anything — especially the sceptical questions.