AI agent memory records remove the sign that a user fact is temporary

Researchers report that AI agents with long-term memory keep chats as short records that can remove the time sign. A record can change I am working at Corvida to The user works at Corvida. This removes the sign that the work can be temporary. The authors name this aspectual flattening and measure it with LAPSE, a benchmark of matched user statements.

How it works

Writer models make the records. Then reader models read them. The words is working show that a fact can become expired, and works does not. LAPSE pairs statements that are different only in the shape of the word.

The numbers

• Three writer models flattened the records in 210 of 381 pairs, and not the opposite.
• The loss was in all 11 models and in three agent memory systems.
• Reader GLM-5.2 acted without a question on 64% of works records, and on 40% of is working records.

What is not known

Gemini 3 Flash showed a smaller distance. One reader that the team added after the test showed the opposite, and others showed none. Some tests were exploratory. Records that kept the words of the user brought back the difference.

Insight: The step that makes memory small can remove meaning

Databricks says its memory changes chats into short facts and gives 3 times more correct results in its own tests. One more study found that compaction removes about 1 third of the rules of a task. Short records seem to save cost for vendors, but the loss shows when an agent acts on an expired fact. Tell vendors to keep the words of the user.