What Should an Agent Forget? Separating What Is Sto... | AI Research

What Should an Agent Forget? Separating What Is Stored from What Is Used
Persistent language agents often struggle to manage memory as they accumulate information over time. When facts change—such as a person switching employers—an agent might accidentally provide outdated information or become confused by conflicting data. This paper introduces RD-Forget, a training-free framework designed to solve this by explicitly separating the information an agent keeps in its long-term archive from the specific evidence it selects to answer a current question. By distinguishing between storage and usage, the agent can retain historical records for future reference while ensuring they do not interfere with present-day accuracy.

How the Framework Works

RD-Forget operates through three primary stages: curation, selective forgetting, and budgeted retrieval. First, a frozen language-model curator examines the agent's stored history and organizes facts into "semantic slots." This allows the system to recognize when a new piece of information supersedes an old one within a specific category, such as a job title or a location. The same ai search question is explored in Trace2Tower, which adds a research perspective.
The framework uses "same-slot replacement" to suppress obsolete facts, preventing them from appearing in the current answer context. However, because the original data remains in the archive, the agent can still access these "forgotten" facts if a user asks a historical question or requires a timeline of events. A rate-distortion formulation then guides the final selection of evidence, ensuring the most relevant information is packed into the agent's limited memory budget.

Managing Conflicting Information

A core challenge for persistent agents is that a fact that is "wrong" for a current query might be "right" for a historical one. RD-Forget addresses this by using intent-aware retrieval. When the system detects that a user is asking about a past state or a chain of events, it can "rescue" superseded entries from the archive. This allows the agent to maintain a clean, up-to-date view for standard questions while preserving the ability to perform complex, multi-hop reasoning that requires older, contextually relevant data. The same ai evaluation question is explored in DRACO, which adds a research perspective.

Performance and Results

The researchers evaluated RD-Forget across several tasks, including conversational memory, knowledge updates, and fact consolidation. Across four different language models, the framework consistently outperformed existing methods. In fact-consolidation tasks, which specifically test an agent's ability to resolve conflicting information, RD-Forget achieved significant gains, with improvements ranging from 11 to 26 percentage points over baseline configurations.
Ablation studies—tests where specific features were removed—revealed that every component of the framework plays a vital role. Configurations that lacked query-conditioned curation or the ability to perform selective forgetting saw the largest drops in performance. These results suggest that the ability to selectively control which facts influence an answer is essential for building reliable, long-term language agents. The ai agents story also surfaces in Stanford Researchers Develop TRACE to Fix..., adding another angle. as detailed in the full paper on Arxiv