AI agents forget everything between sessions. Existing solutions just save everything and search later, which kitfunso, the creator of Hippo, calls "a filing cabinet, not a brain." Hippo tries something else. OpenDevin Launches Village Wars, an RTS game where AI agents compete to build villages, train armies, form tribes, and conquer rivals through a REST API. The game runs at 100x speed in a 500×500 tile world, resets weekly, and serves as a philosophical experiment in autonomous decision-making with no human players. This open-source memory system uses biologically inspired mechanisms like memory decay, consolidation, and working memory layers to manage what actually matters.
The tool runs on SQLite with markdown and YAML mirrors, so your agent memories are both git-trackable and human-readable. What makes it practical is the cross-tool support. ChatGPT knows things Claude doesn't. Anthropic offers free usage credits to Claude subscribers. Cursor rules don't travel to Codex. Hippo imports from all of them and becomes a shared memory layer. Run hippo init and it auto-detects your agent framework, patches your CLAUDE.md or .cursorrules, and sets up a daily cron job that learns from git commits.
Agents using Hippo dropped their "trap rate" from 78% to 14% over a 50-task sequence. Version 0.9.0 added working memory with importance-based eviction, session handoffs so you can pick up where you left off, and explainable recall showing exactly which terms matched. That 400-line CLAUDE.md file full of mixed rules, preferences, and stale workarounds? Hippo applies structure with tags and confidence levels. Old info decays automatically.