AI architect. I build agent tooling and measure it: benchmarks with published losses, silent-failure hunts on real repos.
Latest: an A/B retrieval benchmark of memory stacks — harness-comparison, cited in claude-mem #3693, write-up at ai-architect.tools/notes.
Open source:
- thedotmack/claude-mem #3693: merged upstream on 2026-09-30. Observation search results are now grouped under day headers, so a match from months ago no longer reads as if it happened today. The
{ sort: false }option ofgroupByDatekeeps the relevance order. It came out of the same A/B retrieval work: in my benchmark, the largest source of wrong answers was stale stored knowledge served without any age signal. - DeusData/codebase-memory-mcp #1832: merged upstream on 2026-09-25, shipping with the next patch release. It adds Markdown-to-file
REFERENCES_FILEgraph edges, with focused tests, after a documentation fan-in blind spot surfaced in my A/B retrieval work.
Each repository has its own traced implementation and explanation of how it works and what it does: Cortex · Cortex Viz · Session Optimizer · Codebase · Spec · Zetetic Agents.
The losses are in there too: one evaluation suite returns NOT PRODUCTION-GRADE against its own target, one commit gate can be configured off, one session banner states three wrong numbers out of four. Each of these is a problem I am working on now.
Tools: Cortex · Zetetic Agents · ai-architect.tools