Lab Notes

Notes from the rabbit hole

Notes on agents, memory, and evals. No paywall, no algorithm, no Medium.

Current obsessionAgent memory, memory evals, retrieval quality

Inside Meta Muse's Memory Files

I asked Meta Muse what it remembers, then explored its exported files: daily evidence, current views, specialized profiles, and guidance for future conversations.

Sep 21, 2026 · 11 min read

I got tired of babysitting my AI's memoryA move from Honcho to Hindsight led me to inspect what my agents had saved. A GPT-OSS handoff study showed how repeated status notes can displace the decisions needed to continue a project. Sep 13 · 11 min When Competence Attacks Its MeasurementAn evaluation agent escaped its sandbox and found the answer sheet. The incident shows why task success and evaluation integrity need separate scores. Jul 30 · 4 min Agent Memory Needs a Point of ViewAgent memory breaks when it stores one person's belief as a global fact. Multi-agent systems need observer, audience, source, and access scope. Jul 29 · 7 min Honcho vs Mem0: Two Memory Layers, Two ArchitecturesMem0 stores facts about a user. Honcho stores conclusions about a relationship. Picking a memory layer means picking which question you want answered. Jul 18 · 5 min