Episode Memories: custom eval for episode-level retrieval

[abbudjoe]

Context

Episode Memories (Issue #190) introduces write-time narrative synthesis — detecting episodes across sessions and generating progressive summaries that surface via standard vector search.

Task

As part of Phase 4 (MCP Tools & Polish), create a custom evaluation suite for episode-level retrieval:

Scenarios to test:

1. House-hunting scenario (from design doc) — 5 sessions, cross-session activity, aggregation query ("How many properties before Brookside?")
2. Job search scenario — applications, interviews, offers across sessions
3. Debugging episode — production incident spanning multiple sessions with different services
4. Trip planning — flights, hotels, activities discussed across sessions

Metrics:

• Episode summary surfaces in top-3 recall results for aggregation queries
• Constituent memories are reachable via graph expansion from episode summary
• Queries that don't embed near individual memories DO embed near episode summaries
• Progressive summary accuracy vs full regeneration (drift detection)

Success criteria:

• All scenarios pass with episode summaries surfacing correctly
• No regression on existing LoCoMo/LongMemEval benchmarks
• Latency: remember() adds <10ms synchronous overhead (async detection is unbounded)

Part of

Episode Memories epic (#190), Phase 4