Why versioned memory is non-negotiable for long-running agents

The first thing every memory system gets wrong is treating facts as immutable. They are not. Q3 revenue targets get revised. Drug dosages get updated. A client changes counsel mid-matter. If your memory system overwrites the old value with the new one, you have lost the answer to the most important question in regulated work: what did we believe on date X?

Kalairos stores every mutation as a new version. Each fact carries event time (when it was true), ingest time (when we learned it), a valid interval (the window during which it held), and a source chain. Queries can ask for the current state, or any historical state, with a single asOf parameter.

This is the bitemporal model from data warehousing, ported into agent memory. It is what lets a legal AI cite the law as it stood on the date of the contract. It is what lets a medical agent answer 'what protocol was in effect when this prescription was written.' Without it, every long-running agent quietly loses the truth as it ages.

The API surface is small: remember(fact), query(question), queryAt(question, timestamp), getHistory(entity). Three to four calls cover most agent workflows. Documentation and a 5-minute starter live at krishnalabs.ai/products.