Hippo Memory

Hippo is a CLI-based memory layer for AI coding agents that models human hippocampal memory mechanics: decay by default, retrieval strengthening, sleep consolidation, and reward-proportional decay. It solves the problem of agents forgetting everything between sessions and context being trapped inside individual tools like ChatGPT, Claude, or Cursor. Hippo uses a SQLite backbone with markdown/YAML mirrors, making memories git-trackable and human-readable, with zero runtime dependencies.

• Decay by default — every memory has a configurable half-life (7 days default); persistence must be earned through retrieval or explicit pinning
• Retrieval strengthening — each recall extends a memory's half-life by 2 days, implementing the testing effect mechanically
• Sleep consolidation — hippo sleep compresses repeated episodic memories into semantic patterns, removes decayed entries, and merges related episodes
• Error memories stick — memories tagged --error receive 2× the half-life and a 1.5× strength multiplier, so hard lessons don't fade quietly
• Confidence tiers — memories carry verified, observed, inferred, or stale labels so agents know how much to trust each piece of context
• Hybrid search — BM25 keyword search with optional cosine embedding similarity via @xenova/transformers for semantic recall
• Cross-tool import — import memories from ChatGPT exports, CLAUDE.md, .cursorrules, or any markdown/text file with hippo import
• Auto-hook install — hippo init detects Claude Code, Codex, Cursor, OpenClaw, and OpenCode and patches the right config file automatically
• MCP server — hippo mcp starts an MCP server over stdio, exposing tools like hippo_recall, hippo_remember, and hippo_context to any MCP-compatible client
• Active invalidation — hippo learn --git detects migration commits and actively weakens memories referencing replaced patterns; manual invalidation also supported
• Architectural decisions — hippo decide stores one-off decisions with a 90-day half-life and supports superseding chains as architecture evolves
• Working memory & session handoffs — bounded scratchpad (hippo wm) for current-session notes and hippo handoff for persisting session summaries across agent switches
• Outcome feedback — hippo outcome --good/--bad applies reward-proportional decay modulation inspired by R-STDP in spiking neural networks
• Token budgets — hippo recall --budget 2000 ranks memories by relevance × strength × recency and fills the budget with highest-signal results first