Hippo
A memory layer for AI coding agents that learns from outcomes rather than only storing everything: a memory marked wrong stops being recalled, a newer fact replaces an older one, and used memories get reinforced.
Hippo is a memory layer for AI agents built around learning from outcomes instead of storing and searching everything indiscriminately. When a recalled memory turns out to be wrong, marking it bad removes it from later top results; when a fact changes, the newer version supersedes the old one; and memories that keep getting used are reinforced, so a coding agent stops repeating a mistake it has already been corrected on.
It runs on SQLite with no other runtime dependencies, works over MCP or HTTP with any command-line agent, and keeps CLAUDE.md notes, Cursor rules, ChatGPT exports and Slack history queryable from one store instead of being locked to a single tool. Its authors publish benchmark results and pre-registrations for their retrieval claims, including at least one earlier result that was later retracted, alongside an optional hosted reranker.
It installs with a single npm command that scans a machine for git repositories and adds memory to each one.