SurrealDB Agent Memory splits valid, known and system time

A practical mismatch is the starting point: when old conversations or documents are imported into an AI's memory later, the date of the event and the date the AI learned about it can drift apart. An example makes this concrete — a log saying someone has worked at a company since April, imported in July, yields three distinct moments: April for validity, July for when the AI learned it, and the actual database write time for storage. Collapsing them into one date makes a question like 'did the AI know about this job in May?' hard to answer.

The piece also points to a real implementation discussion. A Graphiti developer reported Issue #1489 in May 2026 asking to separate event time from import time during backfill of old conversations, though the issue was later marked resolved upstream in July — so it is not evidence of a current defect. The takeaway is narrower: how timestamps are passed when importing past logs can become a design question.

The SurrealDB material is presented as that product's own design explanation. Its three-temporal model uses valid_from and valid_until for the validity period and as_of to trace what the system knew at a given point, while observed_at lets a caller supply a known time different from the import time. This is not a standard feature of every AI memory, nor a guarantee that dates are read correctly automatically — and keeping history is separate from being able to automatically correct a wrong memory.