PMB: Stop re-explaining your project to AI coding agents

You’re hitting on two different things – isolation and staleness – so let me break them down.

Cross-project leakage is controlled by isolation, not by ranking. Each project/client is its own workspace with a separate store: its own SQLite events, vector indexes, BM25 indexes, and graphs under ~/.pmb/workspaces//. The scope of recall is up to the active workspace, so one client’s memory may not score with reference to another. Leakage between customers is not a ranking issue here – it has been closed.

The keyless reversal decision is the honest difference, and you said it absolutely right: an architectural direction that flipped mid-project, no clean key, no obvious “we switched from X to Y.” There is no semantic conflict-detection today, so the old decision keeps scoring on BM25+ vectors until recency/forgetting-curve decay reduces it, an improvement overrides it, or someone archives it. The dashboard shows the timeline and both decisions, but it doesn’t currently auto-flag that there is a conflict between the two – so worst case, yes, the stale one scores good as long as it’s not manually archived. I won’t pretend otherwise.

This is actually the next build: reversal/conflict detection that connects a new decision to the decision it deleted even without the clean key, plus a needs-review surface in the dashboard that marks candidate conflicts for one-click supersede/archive, so it’s not human noticeable. This is the difference between trusting it and constantly auditing it for a multi-client setup like yours, which is why it’s high on the list.

Repo if you want to follow where it leads: https://github.com/oleksiijko/pmb