DMF: A Deterministic Memory Framework for Conversational AI Agents

Researchers have presented two new frameworks aimed at improving AI agent performance and safety: DMF, a deterministic memory framework for conversational AI, and a new framework for establishing refusal boundaries in AI agents for cybersecurity tasks.

The Deterministic Memory Framework (DMF) eliminates the need for large language model (LLM) calls in memory management, reducing token costs to nearly zero^[1]. DMF uses a CPU-first approach, assigning each conversational interaction a Survival Score computed from deterministic content signals and conversational cues. In experiments, DMF achieved comparable accuracy to Mem0, a popular memory layer, while using zero tokens to prepare the memory context and 5x to 242x fewer tokens over the entire conversation^[1]. Meanwhile, researchers have proposed a new framework for establishing refusal boundaries in AI agents for cybersecurity tasks, defining criteria for task refusal and an evaluation methodology for measuring agent robustness^[2]. The framework aims to address the increased risks in cybersecurity resulting from improved LLM performance on complex tasks. Only 2 out of 8 frontier models tested showed meaningful refusal behavior under the new framework^[2].