Hey Product Hunt 👋 I built CogniKernel because I kept having the same conversation with Claude Code. Session ends, I come back the next day, and it wants to re-litigate a decision we settled last week — or worse, it retries the exact approach we already found doesn't work. I was hand-maintaining CLAUDE.md like a changelog and still losing context. Every memory tool I tried solved this by shipping my session transcripts to an LLM to "summarize what mattered." That means another API key, per-session token cost, and my code leaving my machine. That tradeoff didn't sit right. So CogniKernel does it differently: extraction is *classification, not generation*. Two small fine-tuned encoder models (~130MB ONNX) run locally on CPU in milliseconds — one scores whether a sentence is worth remembering and what type it is, the other detects when a new decision supersedes an old one. No API calls, nothing leaves your machine. The only LLM involved is the coding agent you already run. A few things I'm proud of: - Abandoned approaches go in a graveyard, so the agent doesn't re-try dead ends - A "don't do X" rule surfaces at the moment you're about to do X — not three files later - One store shared between Claude Code and Codex in the same directory - Every hook is fail-open: memory can break without ever breaking your session Benchmarked against flat notes and no-memory across four multi-session projects, it made 2–4× fewer file reads in every single one. It's Apache-2.0 and on PyPI: pip install "cognikernel[embedding]" I'd love feedback from anyone running long-lived projects through Claude Code — especially on what memory *should* be captured that I'm currently missing.

CogniKernel

Give Claude Code and Codex memory that survives sessions

Give Claude Code and Codex memory that survives sessions