Home / Blog / Bitemporal AI memory
Bitemporal AI memory preserves what the agent knew then.
Ordinary memory stores the latest version of a fact. Bitemporal memory preserves when the fact was true and when the system learned it, so historical decisions can be reproduced without future knowledge.
Definition
What is bitemporal AI memory?
Bitemporal AI memory records two independent timelines for every fact. Event time describes when the fact was valid in the real world. System time describes when the AI system received, revised, or retired that fact. Together they let an agent answer current questions while also reconstructing the exact information available during a prior decision.
The two clocks
Event time and system time answer different questions.
When did this fact apply in the world? A contract amendment may become effective on January 1 even if it arrives later.
When did the agent learn this fact? If the amendment was ingested on January 12, decisions made on January 5 could not have used it.
Each version receives start and end boundaries on both timelines instead of overwriting the previous record.
Why it matters
Single-timeline memory silently rewrites history.
Suppose a customer risk rating is corrected from low to high. A normal database updates the row. A later audit then sees the high rating and may incorrectly conclude that an earlier decision ignored it. Bitemporal memory preserves the original value, the correction, its effective date, and the moment the system received the correction.
This prevents lookahead bias in evaluations, backtests, compliance reviews, investigations, and any workflow where facts or policies change after an agent acts.
Query model
A point-in-time query applies both temporal boundaries.
Return the version valid now using the latest accepted system record.
Return only versions that had entered the system by the selected historical timestamp.
Ask what is now believed to have been true on a past date, including corrections learned later.
Implementation checklist
How to evaluate a bitemporal memory layer.
Corrections should close validity windows, not destroy prior records.
Backdated information ingested later must not appear in earlier decision replays.
Each result should retain source identity, provenance, and the temporal window that made it eligible.
The same as-of query should produce the same eligible memory set and evidence package.