An AI memory layer that forgets on purpose

What it does

Hippo is a long-term memory layer for CLI AI agents. It stores observations, errors, and context in a local SQLite database, then surfaces the right memories when you switch between tools like Claude Code, Cursor, or Codex. A nightly consolidation pass prunes stale entries, merges duplicates, and promotes important lessons to a global store so they survive across projects.

The interesting bit

Instead of hoarding every prompt like a semantic search junk drawer, Hippo copies actual neuroscience: memories decay by default, retrieval strengthens them, and tagged errors fade more slowly than casual observations. The result is a portable, provenance-tracked memory that aims to stop you from hitting the same deploy bug every week. Even the benchmarks are unusually honest—one claimed improvement was formally retracted in the changelog, and negative results for the opt-in Jev reranker are published alongside the positive ones.

Key highlights

• Zero runtime dependencies; runs fully offline with an optional API-based reranker.
• Imports and exports across ChatGPT, Claude (CLAUDE.md), Cursor (.cursorrules), Slack, and markdown.
• Every memory carries provenance (kind, scope, owner, artifact_ref) and supports GDPR-correct deletion.
• Multi-tenant with scrypt-hashed API keys and negative tests proving tenant isolation.
• 926 tests run against a real database; the project explicitly bans mocks.

Caveats

• The opt-in Jev reranker improves ranking recall but, by the project’s own admission, did not beat the free local cross-encoder on graded answer quality in three tests.
• Connectors for GitHub, Jira, and Notion are listed as “next” after Slack, so broader system-of-record ingestion is still pending.

Verdict

If you juggle multiple AI agents and want a single, portable memory that learns from your mistakes, Hippo is worth the SQLite footprint. If you just need a simple vector store for RAG, the biological decay model and multi-layer consolidation will feel like overkill.