I agree that similarity search can't surface reliable project judgment alone. How do you know what was true, what was rejected, and what's obsolete since changes in the project? I am actually building Open-Latch for this reason. It leverages project decisions, facts, lifecycle state, reconciliation, enforcing cited evidence and a judgement gate before coding agents act.

GitHub: https://github.com/open-latch/latch

I felt the same annoyance as OP mentioned, but focused on agents violating project judgement and direction. Obviously we are optimizing for some different things than deja: deja uses no-LLMs and exact verbatim transcript retrieval.

Latch uses LLMs, ranking, etc, to emphasize strong judgment, correction, and enforcement over dynamic project flow.