Hippo Memory

Hippo Memory gives coding agents persistent memory across projects, sessions and tool changes. It's for developers who want agents to retain useful lessons and stop recalling facts that have proved wrong.

• Shared agent memory: Claude Code, Codex, Cursor, OpenClaw, OpenCode and Pi can use the same store, and other MCP clients can connect.
• Corrections and replacement: Mark recalled memories wrong to lower their ranking, replace outdated facts to remove the old version from recall, or reject a memory entirely.
• Session continuity: Capture failed Claude Code tool calls as error lessons and preserve task snapshots across session compaction; frequently recalled memories last longer.
• Existing notes and history: Import ChatGPT exports, CLAUDE.md, Cursor rules, Slack history and Markdown, with separate project stores and a global store for shared lessons.

Hippo is open source under MIT and runs on macOS, Windows and Linux. It stores memories locally in SQLite, with Markdown copies you can inspect and commit. Default search needs no model, account or network connection. Optional local MiniLM embeddings add semantic search.

Some features send memory text to external providers. Consolidation uses Anthropic for fact extraction whenever an Anthropic API key is present, unless that extraction is disabled. API embedders and hosted rerankers also send text outside your machine.

Published evaluations measure retrieval and suppression of incorrect memories; they don't establish improved performance on agent tasks.