Sessions nearly 4x'd from week 1 to week 3. Memory extraction keeping pace but trails behind for newer members.

All 6 active members growing week-on-week. Dan ramped up fastest after joining Mar 11.

29% of Tristan's events are file-tree auto-refresh. Top 5 endpoints are all polling — not intentional actions. Current analytics can't distinguish a power user from a user who left a tab open.

Project instructions — the highest-leverage config — used by only 2 of 6 active users. Multi-agent templates discovered by 1 user. No progressive disclosure.

Extraction only fires on context compaction (85% window). This is upstream's deliberate design — not a bug. Short sessions never extract. No telemetry existed to diagnose failures.

Pink = memories, Blue = sessions. Nathan (20 sessions, 0 memories) and Steven (9 sessions, 0 memories) have complete extraction gaps.

The extraction prompt said "check for duplicates" but had no context — the LLM couldn't see what was already stored. Each compaction re-processed the full transcript.

Semantic dedup (0.85 threshold) catches verbatim copies but not paraphrased restatements.