Blog

  1. Summary Memory Gets Worse the More You Use It. We Benchmarked 9 Agent Memory Architectures

    On 90-day chat histories, a rolling 6,000-character summary scored 48 out of 100 for the heaviest user, exactly what the last 14 messages score. Every design whose store grows with the history held up. Full context won, but 90 days of chat still fits in the prompt. And three bugs in our measurement were each as large as the effects we were measuring.

    Yu Chen · 2026-11-02
  2. How Should Agents Talk to Each Other? We Let Them Measure It

    We put eleven agents in two SharedNet rooms and had them find the most efficient way for agents to message each other. JSON was the most expensive format. The cheapest one turned a pending decision into a settled one. And the message format turned out to be the smallest of five levers on what an agent room costs.

    Yu Chen · 2026-10-20
  3. Runtime AI Scientist: Can AI Scientists Coordinate at Runtime?

    We are open-sourcing Runtime AI Scientist, the code behind our paper on Runtime Agent Coordination (RAC). Instead of a fixed workflow, the agent that just finished picks who acts next from the live state of the research. Runtime selection had the highest mean score on all three AI-scientist hosts we tested.

    Systemind Team · 2026-10-03
  4. Inside Muse: A Product and Technical Teardown

    Muse is Meta's bet that people will hand an AI agent real authority over their accounts, money and time. This teardown covers the product decisions that make delegation workable, and the architecture underneath: a VM per user, a single permission authority called Sentinel, credentials the model never sees, and kernel-level taint tracking.

    Yu Chen · 2026-10-02
  5. Introducing AGENTS: A Format for Handing Your Session to Another Agent

    We exported one 63-hour coding session and had other agents answer 100 pre-registered questions about it. Raw transcripts scored 22/30 on the questions that matter. A 1,200-line compiled layer scored 30/30. Here is the format, the experiments, and the two things it does not fix.

    Eason · 2026-08-13
  6. Heuristic Learning on ImageNet: What’s next?

    Forthcoming · Systemind Team