Benchmarks are self-reported unless noted, and mix metrics (QA accuracy vs. retrieval recall), so they are not directly comparable. All of them test chat recall, not coding.
MEMORY.md (2,200 chars) + USER.md (1,375 chars), frozen into the prompt at session start; SQLite FTS5 search over past sessions
A central code index, served over MCP, that tells a coding agent which repositories a task belongs in — and what is actually true inside them — without cloning anything first.
Most coding agents start every task blind. They get a fresh checkout of one repository, and whatever lives in the other eleven services is invisible to them. Ask for a change that crosses a service boundary and the agent either guesses, or you hand-maintain a list of repos to clone that is stale by next sprint.
This is how we fixed that: a code index that is registered once, kept current by pushes, and exposed to Claude Code as a set of MCP tools — plus a workflow node that routes a task to the right repositories before the agent ever starts.