Shadows proves memory isn't enough

This is a sharp reminder that “just add memory” is not a full agent strategy. Retrieval can be near-perfect and the system can still fail at judgment, which is where most production pain actually lives.

• –LongMemEval numbers are strong on recall_all@5, but the overall score stays meaningfully lower because multi-session aggregation and preference handling are still weak
• –The post’s better idea is classifying the query before retrieval, so “latest,” “earliest,” and “all history” use different memory behavior
• –The shared-memory plus per-agent-memory setup pushes the product toward portable identity, not just sticky context
• –The orchestration layer matters as much as the memory layer: agents need to route work, reject bad fits, and ask peers directly instead of bouncing everything through a manager
• –The benchmark framing is useful, but the bigger takeaway is that agent evals need to measure decision quality, not just whether the right fact was retrieved