llms.txt@website/public git:20260925.02d79c1A

Immutable. This exact content never changes and is served at /api/v1/blob/bb184d2456be901d.

# hippo

> A memory layer for AI agents that learns what is wrong and stops repeating it: a memory marked bad drops out of recall, a newer fact supersedes the old one, and retrieval strengthens what gets used. Memories also decay on a half-life (365 days, not tuned) and a sleep cycle consolidates repeats; neither has been shown to improve recall. Open source (MIT), zero runtime dependencies, runs locally.

hippo (npm package `hippo-memory`) is an open-source memory system for AI agents and LLM applications. Unlike "save everything and search later" stores, hippo models the full memory lifecycle: decay by default, retrieval strengthening, outcome-weighted half-lives, error tagging, conflict detection, supersession, and sleep consolidation. It is local-first (SQLite source of truth with git-friendly markdown mirrors, no outbound HTTP unless you opt in to a network reranker or an API embedder), framework-agnostic, and ships an MCP server plus auto-installed hooks for Claude Code, Codex, Cursor, OpenClaw, and OpenCode.

## Key facts
- License: MIT. Runtime: Node.js 22.16+. Zero runtime dependencies (embeddings are optional).
- Storage: local SQLite source of truth plus markdown mirrors. No cloud, no account, no telemetry.
- Retrieval: BM25 out of the box; optional hybrid BM25 + embeddings via a pluggable provider (local MiniLM default, or opt-in OpenAI / Voyage / Cohere).
- Mechanics: decay by default, retrieval strengthening, reward-proportional half-lives, error and emotional tagging, sleep consolidation, conflict detection, bi-temporal supersession.
- Integrations: MCP server (Cursor, Windsurf, Cline, Claude Desktop) and hooks for Claude Code, Codex, Cursor, OpenClaw, OpenCode.

## Benchmarks
- LongMemEval-S, standard per-question haystack, recall@5: 98.0% with the free local MiniLM embedder (an optional install; best of five settings in the benchmark scripts), 99.8% with voyage-3-large (opt-in, measured June 2026), level with the published frontier (gbrain reports 97.6%). Zero-dependency BM25-only path: 74%.
- 3,500+ tests against a real database. No module mocks and no mocked store; only paid network calls are stubbed.

## Links
- Repository: https://github.com/kitfunso/hippo-memory
- npm: https://www.npmjs.com/package/hippo-memory
- Docs (README): https://github.com/kitfunso/hippo-memory#readme
- Benchmarks: https://github.com/kitfunso/hippo-memory/tree/master/benchmarks
- Changelog: https://github.com/kitfunso/hippo-memory/blob/master/CHANGELOG.md

## Install
```
npm install -g hippo-memory
hippo init --scan ~
```

Agents installing hippo: follow https://github.com/kitfunso/hippo-memory/blob/master/llms-install.md (install, wire in, verify with `hippo doctor --json`). Zero-install MCP: `npx -y hippo-memory mcp`.