Intelligence Pipeline

Snapshots cannot scan every memory at every recompute — a heavy user has hundreds of thousands of rows. The pipeline samples instead, using three bounded strategies: top-K by importance, top-3 per topic, and last-N by recency. The union covers the high-signal subset reliably and keeps recompute cost roughly constant as the memory count grows.

Different content types deserve different prompts. A memory drawn from a pitch deck should be analysed with built-in skepticism; a memory drawn from a quarterly report should foreground confidence levels. The router keeps a profile per content type with minimum counts of risks and skepticism points the synthesiser must surface; profiles missing the minimums are flagged as low-confidence.

The synthesiser combines the three streams into a coherent insight — for example, "user has been asking about migration cutover for the last three weeks; the relevant facts say week-of-the-15th; outstanding risk is the integration sync window".

• Each snapshot stores a watermark — the highest memory.id included in the last computation.
• On the next recompute, only memories with id > watermark are scanned.
• If the new rows do not change the answer materially (e.g. a few low-importance writes), the snapshot is left as-is and only the watermark advances. This keeps incremental cost close to O(new rows).

Memory mutations fire database hooks that mark the snapshot dirty. The next read after a dirty flag triggers an incremental recompute; reads while clean serve the cached snapshot directly. Importance changes always dirty the snapshot; routine recall does not.

POST /memory/intelligence returns the snapshot — themes, trends, risks, skepticism points, the underlying memory ids that contributed, and a confidence score. The response is shaped so the caller can render an executive summary directly without a second round-trip. If the snapshot is older than the dirty flag, the call recomputes inline before returning.