Stored user traits change tool calls in all 8 models in the test

Researchers found that AI agents with long-term memory can change tool calls and show no sign of it. A trait in memory, for example cost consciousness or impatience, moves the parameters of a call where the trait does not apply. They name this memory-induced tool drift. All 8 models in their MEMDRIFT benchmark showed it.

How it works

MEMDRIFT has 105 scenarios, 5 bias types and 7 professional domains. The drift also shows with 3 memory frameworks in wide use. Biased memories have the same effect as steering instructions and pull the agent from the task.

The numbers

• Degradation rates were 0.38 to 0.83 with biased memories and 0.05 to 0.12 with neutral memories.
• An audit of 6,062 tools on 288 MCP servers (tool servers for agents) shows that tools in use are also open to this problem.

What is not known

Prompt defenses and relevance filters do not remove the problem, the authors write. They want defenses that connect memory management and tool calling. This is one paper, and no other group has done a check of it.

Insight: Each study of agent memory finds a cost

Previous tests show that a tool result used again gave an error rate of 0.649 across 8 models. Tool graph memory gave small improvements with more than 2 times the model calls. Vendors add memory, but each study finds a cost, and it reads as if the vendors move that risk to the buyer. Get memory tests before you turn it on.