Memory Session Search

On-demand recall over the CALLER'S OWN past sessions (trajectories) — the recall half of two-tier memory (Hermes-style): curated memory is always in context; this searches "what happened in past sessions".

Plan-gated (Plus/Pro). community has a 0 monthly quota and is short-circuited to a self-describing gated payload BEFORE any disk scan, so the free tier can never trigger the per-call summary+raw file scan. Paid tiers are bounded three ways: the SAME visibility window as trajectories (you can only recall what you can still see), per-plan scan caps (max_sessions + raw-fallback budget), and a monthly call quota.

Summary-first with a bounded raw-JSONL fallback for unsummarized sessions (see app/session_search.py), so recall is complete without a global FTS index or a forced re-summarize. Reads only the caller's files; the heavy IO runs in a thread. Degrades to an empty result on error.

Contract v2 body: {query, scope?, project_ids?, session_ids?, context_session_id?, limit?, before?, after?}. context_session_id is supplied by Core from ToolContext and is never a model-facing argument.