kitfunso/hippo-memory

What it solves

Hippo is a memory layer for AI agents that prevents them from repeating mistakes and losing context when switching between different AI tools (like Claude Code, Cursor, and Codex). Unlike traditional semantic search systems that save everything, Hippo implements a biological model of memory where information decays over time unless it is reinforced through use, ensuring that agents retain only the high-value lessons and patterns.

How it works

Hippo uses a SQLite backbone with markdown mirrors and operates across three biological layers: a session-only buffer, an episodic store for timestamped memories, and a semantic store for compressed patterns.

• Decay and Reinforcement: Memories have a default half-life (e.g., 7 days). Recalling a memory extends its half-life, while tagging a memory as an error makes it decay slower.
• Sleep Consolidation: The hippo sleep command merges related episodic memories into stable semantic patterns and prunes decayed information.
• Integration: It integrates with agent frameworks via hooks (e.g., patching CLAUDE.md or .cursorrules) and can import data from ChatGPT, Slack, and git commit histories.
• Retrieval: Memories are ranked by relevance, strength, and recency, and can be constrained by a token budget to avoid context stuffing.

Who it’s for

Developers using multiple AI coding agents and CLI tools who want a persistent, tool-agnostic memory store that learns from their project history and prevents repetitive errors.

Highlights

• Biological Memory Model: Implements decay, reinforcement through retrieval, and sleep-based consolidation.
• Tool Agnostic: Works across Claude Code, Cursor, Codex, and any MCP client.
• Git-Integrated Learning: Automatically extracts lessons from fix, revert, and bug commits in git repositories.
• Provenance and Governance: Every memory tracks its source (kind, scope, owner) and supports GDPR-compliant deletion.
• Confidence Tiers: Labels memories as verified, observed, inferred, or stale to help agents gauge trust.
• Outcome Feedback: Allows users to mark recalled memories as helpful or unhelpful to adjust their decay rates.