October 2026 Summaries

Long-lived consumer AI agents need to distinguish when a fact occurred from when it was stored, a bitemporal approach intended to prevent relative expressions such as “yesterday” or “last weekend” from becoming inaccurate over time. Using a fitness-coach example, the discussion describes Mem0 features for supplying original event timestamps during imports and check-ins, resolving relative dates with appropriate precision and timezone awareness, retaining superseded facts while retrieving only current ones when needed, and setting expiration dates for time-limited information without deleting historical records. It also covers date-based filtering for fixed time windows, automatic temporal ranking for natural-language queries, configurable reference dates for reproducible tests and user-local scheduled tasks, and memory decay that lowers the ranking of rarely used older facts without hiding them. Finally, it explains how historical retrieval and per-memory change histories can help determine what an agent knew at a particular time, supporting audits of past advice or user complaints.