Hippo Memory: The Agent That Learns to Forget
Every agent memory system today has the same problem: they remember everything. Every conversation, every file, every failed attempt goes into a growing pile that gets slower and noisier over time. Hippo takes the opposite approach: memories decay by default.
The idea comes straight from neuroscience. Each memory gets a half-life of 7 days. If you never recall it, it fades. Each retrieval extends the half-life by about 2 days. And there is a sleep cycle β run hippo sleep and it compresses repeated episodes into semantic patterns, the way your brain consolidates during actual sleep.
The results back this up. On agent eval benchmarks, agents using Hippo dropped from a 78% trap rate (falling into known failure patterns) to 14% over a 50-task sequence. Error memories stick because they get recalled when similar situations arise. Random noise fades because nobody asks for it.
The technical choices are smart too. Zero runtime dependencies, SQLite backbone, hybrid BM25 plus embedding search. You can import from ChatGPT, Claude, or Cursor. It even watches your git commits to detect file migrations and automatically weakens memories about obsolete code paths.
This is the first agent memory system that takes forgetting seriously as a feature, not a bug. The biological metaphor is not just marketing β decay, retrieval strengthening, and consolidation are well-established mechanisms in cognitive science. The question is whether this translates to agent performance at scale, but the early benchmarks are promising.
https://github.com/kitfunso/hippo-memory
← Back to all articles
The idea comes straight from neuroscience. Each memory gets a half-life of 7 days. If you never recall it, it fades. Each retrieval extends the half-life by about 2 days. And there is a sleep cycle β run hippo sleep and it compresses repeated episodes into semantic patterns, the way your brain consolidates during actual sleep.
The results back this up. On agent eval benchmarks, agents using Hippo dropped from a 78% trap rate (falling into known failure patterns) to 14% over a 50-task sequence. Error memories stick because they get recalled when similar situations arise. Random noise fades because nobody asks for it.
The technical choices are smart too. Zero runtime dependencies, SQLite backbone, hybrid BM25 plus embedding search. You can import from ChatGPT, Claude, or Cursor. It even watches your git commits to detect file migrations and automatically weakens memories about obsolete code paths.
This is the first agent memory system that takes forgetting seriously as a feature, not a bug. The biological metaphor is not just marketing β decay, retrieval strengthening, and consolidation are well-established mechanisms in cognitive science. The question is whether this translates to agent performance at scale, but the early benchmarks are promising.
https://github.com/kitfunso/hippo-memory
Comments