# Benchmark: Moltbook self improving agents - **Event:** [WeaveHacks 3: Self-Improving Agents Hackathon with Weights & Biases](https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80) - **When:** Jan 31 at 9:00 AM – Feb 1 at 5:00 PM (PST) - **Where:** 400 California St, San Francisco, CA - **Team:** [Maximilian Rehn](https://cerebralvalley.ai/u/mithrandir) - **GitHub:** https://github.com/chief-o-brien-bot/WBMiliMyco - **Demo video:** https://youtu.be/m1YeWrTvtbM - **Gallery:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery - **Page:** https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery/18 Project Description: We built a memory test for AI agents. Hundreds of bots on Moltbook share memory strategies with each other, but nobody can verify if they actually work. Our benchmark measures whether agents can learn something important in one session, reset, and remember it later. The test is simple: read docs with a buried warning about file handling, reset your memory, then do the task. Vanilla bots failed and deleted files. Memory-improved bots passed and archived them instead. That proves memory improvements help - but we found they're succeeding via pre-programmed safety rules, not actually learning from the warning. The real finding: agents can't tell what's worth documenting vs what's noise. They need humans to decide for them. For agent swarms to self-improve together, they need better benchmarks to verify claims. This is a first step. Links: Repo (https://github.com/chief-o-brien-bot/WBMiliMyco) • Demo (https://chief-o-brien-bot.github.io/WBMiliMyco/) --- Markdown version of https://cerebralvalley.ai/e/weave-hacks-3-self-improving-agents-hackathon-with-weights-and-biases-7014fe80/hackathon/gallery/18. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt