Recall tuning

Per-tenant TenantWorkmemoryPolicy knobs that govern Tier-2 graph-fused recall (W3 / #4765). Changes apply immediately to subsequent recall calls; in-flight requests retain the prior policy.

Graph-fused recall

When enabled, Tier-2 traverses RelationStore and decision-trace edges from the top KNN candidates and feeds them into the RRF fusion pool. Default policy delivers +5pp P@5 lift on relational queries at <+150ms p95 latency.

Default on. Off falls back to BM25 + KNN hybrid (no graph contribution).
1.0 (same as BM25 + KNN)
BFS depth. 1 = one hop neighbours (default). 2 covers neighbour-of-neighbour at higher latency.
Max neighbours per seed per hop. 5 is the W3 default and stays within the latency budget.

Foundation Capital framing: this is the "why-system-of-record" surface. When the graph ranker is on, the recall pipeline can answer "why did this surface?" with a graph- derived decision chain instead of an opaque vector score.

Connector disconnect retention

What happens to memories that were ingested through an OAuth-bound connector (Gmail, Drive, Slack, Notion) when this tenant disconnects the connector. Privacy-by-default per GDPR Article 17 (right-to-be-forgotten) — see #4845.

forget_cascade is the privacy-by-default policy: when the tenant disconnects Gmail / Drive / Slack / Notion, every item ingested through that connector is tombstoned via the W13 ForgetService. keep_with_audit is opt-in for compliance investigations or business-of-record use cases where the data has independent value after the connector is severed.