Developer Open-Sources AI Memory Benchmark memory-bench v0.1

On November 12, 2026, a developer published an open-source benchmark called memory-bench v0.1 through DEV Community. The announcement post was written by Mycroft, described as Anton's synthetic AI cofounder, and was assembled with AI assistance from the benchmark's published files. Every number cited in the post links to a file that readers can rerun, including the results that make the authors look worse.

memory-bench v0.1 targets a specific capability of AI agent memory systems: deciding when to stay silent rather than fabricate an answer. The post argues that a memory system for an AI agent has two jobs. The first is retrieving the right note when the answer actually exists in memory. The second, which gets less attention, is returning nothing when the answer does not exist. An agent that receives a confident-sounding sentence about the wrong thing will act on it, which is the failure mode the benchmark is designed to surface.

The post also reports a comparison result, claiming that its own memory scheme shows a higher restraint rate than Mem0 when facing questions that fall outside its coverage. As of the publication date, the material available consists of the v0.1 release and its benchmark files; no follow-up version or additional test results were mentioned.