Hippo v0.21.0: Biologically-Inspired Memory for AI Agents with Multi-Tool Support

Hippo is a biologically-inspired memory system for AI agents that addresses the problem of agents forgetting everything between sessions. Unlike solutions that just save everything, Hippo implements decay, retrieval strengthening, and consolidation mechanics similar to human memory.
Core Features
Works with: Claude Code, Codex, Cursor, OpenClaw, OpenCode, Pi, and any CLI agent. Imports from: ChatGPT, Claude (CLAUDE.md), Cursor (.cursorrules), and any markdown. Storage uses SQLite backbone with markdown/YAML mirrors that are git-trackable and human-readable. Zero runtime dependencies, requiring only Node.js 22.5+. Optional embeddings via @xenova/transformers.
Setup and Usage
Install with npm install -g hippo-memory. For single project setup: hippo init. For all projects at once: hippo init --scan ~ which finds every git repo under your home directory, creates a .hippo/ store in each, and seeds it with lessons from commit history.
After setup, hippo sleep runs automatically at session end via agent hooks and performs five operations:
- Learns from today's git commits
- Imports new entries from Claude Code MEMORY.md files
- Consolidates memories (decay, merge, prune)
- Deduplicates near-identical memories (keeps the stronger copy)
- Shares high-value lessons to the global store so they surface in every project
Manual usage includes: hippo remember "FRED cache silently dropped the tips_10y series" --tag error and hippo recall "data pipeline issues" --budget 2000.
What's New in v0.21.0
hippo setup — one command for every tool. Detects Claude Code, OpenCode, OpenClaw, Codex, Cursor, and Pi on your machine and installs all available SessionEnd + SessionStart hooks in one pass. Idempotent and safe to re-run.
OpenCode hooks: SessionEnd + SessionStart install into ~/.config/opencode/opencode.json (OpenCode added Claude-Code-compatible hooks in Jan 2026). You now see consolidation output — new SessionStart hook prints the previous session's hippo sleep output between banners on next startup.
Recent Updates
v0.20.3 added visible confirmation on session-end sleep. The installed SessionEnd hook now echoes [hippo] consolidating memory... before running and [hippo] sleep complete or [hippo] sleep failed after. Existing installs need reinstall (hippo hook uninstall claude-code && hippo hook install claude-code) to pick up the new command.
v0.20.2 changed Claude Code hook to use SessionEnd instead of Stop. Earlier versions ran hippo sleep after every assistant turn; now it runs once at session exit. Re-running hippo hook install claude-code migrates existing Stop entries automatically. Hook installation and initialization no longer create accidental CLAUDE.md files — they only patch agent-instruction files that already exist.
📖 Read the full source: HN LLM Tools
👀 See Also

SkillMesh: MCP-Friendly Router for Large Tool Catalogs Reduces Context Size by 70%
SkillMesh is an MCP-friendly router that retrieves only relevant expert cards for AI agent queries, reducing context size by 70% and improving tool selection. It supports Claude via MCP server, Codex skill bundles, and OpenAI-style function schemas.

Claude Code Best Practice Repo Hits 50k Stars, Built Entirely with AI Agents
A GitHub repository packed with Claude best practices, 100% developed and maintained by autonomous Claude code workflows, crossed 50,000 stars — making it Pakistan's most-starred repo in 2026.

Open-source pipeline turns Claude Code workflow into reusable skills
A developer who used Claude Code daily for 9 months has open-sourced a pipeline that structures feature development with checkpoints like functional documentation, technical documentation, complexity estimation, and security checks. The pipeline includes /new-feature and /bug-fix entry points that guide implementation.

yoyo: Local MCP Server for Grounded Codebase Reads and Guarded Writes with Claude Code
yoyo is an open-source local MCP server that provides coding agents like Claude Code with grounded repository reads and guarded writes across 16 languages, including Rust, Go, Python, and TypeScript. It prevents broken edits from silently landing by returning machine-readable guard_failure output and enabling retry_plan for targeted repairs.