Inspiration

A semantic memory layer for incident response agents — compresses past postmortems into token-tiny chunks, generates actionable artifacts, and blocks risky PRs before they ship.

Problem I ran into

Deduplication across sessions. ChromaDB stores vectors by ID, and our first implementation generated a new UUID every time a postmortem was parsed.

Accomplishments that I'm proud of

Concrete, measurable token compression. Six postmortems go from ~12,000 tokens to ~900 — a 13× reduction. That's not a marketing number; it's a direct consequence of the schema design. Each chunk stores only what's needed for retrieval and synthesis, and the math shows up live in the dashboard.

Built With

Share this project:

Updates