Memory Extraction — Dedup Context & Telemetry

What changed

Two improvements to the memory extraction pipeline shipped on 19 March 2026.

1. Existing memories fed into extraction prompt

Before calling smartExtractMemories(), the system now fetches up to 10 existing memories for the project and includes them as “Already Known” context in the extraction prompt. The LLM instruction was updated to: “Do NOT extract anything already covered by the Already Known memories — even if worded differently.”

This prevents paraphrased duplicates. For example, one user had 3 near-identical memories about PaaS classification — each extracted during a separate compaction with slightly different wording. The semantic dedup (0.85 threshold) didn’t catch them because the titles and embeddings were different enough.

2. Extraction telemetry

Structured logging added at two points:

• smartExtractMemories() — logs extraction_skipped with reason (transcript_too_short or no_api_key) when extraction is bypassed
• deepExtractMemories() — logs extraction_complete with project, user, transcript length, candidates from LLM, stored vs skipped counts, skip reasons, and memory types

Grep production logs for [Memory:telemetry] to diagnose extraction gaps.

3. “For Review” dashboard banner

The unextracted memories card was moved from a buried position in the dashboard grid to a full-width amber banner between the Hero KPIs and the grid. Each chat now has both Extract and Ignore buttons.

Files changed

File	Change
memory-mcp/src/chat.ts	~50 lines added — fetchExistingMemoriesForPrompt(), buildExtractionPrompt(), telemetry logging
admin-ui/src/features/dashboard/UserDashboardPage.tsx	For Review banner wrapping UnextractedMemoriesCard
admin-ui/src/features/dashboard/UnextractedMemoriesCard.tsx	Review & Extract toggle, Ignore button