RD-Forget is a training-free framework for persistent language agents. It keeps all past observations in an archive but builds a query-conditioned memory view for each answer. A frozen language model curator identifies relevant facts, groups them into semantic slots, and preserves relations needed for multi-hop reasoning. Superseded facts are suppressed for current-state questions but remain eligible for historical queries via intent-aware retrieval. The system uses a rate-distortion formulation to pack evidence within a memory budget. Experiments across conversational memory, knowledge updating, fact consolidation, long-context reasoning, and personalization show that forgetting and query conditioning are critical. Configurations without them have the largest performance deficits. Slot grouping, historical access, and relation preservation each contribute complementary functions.
Agents can retain old facts without using them indiscriminately
RD-Forget preserves an agent’s full observation history while constructing a query-specific memory view that suppresses obsolete facts when appropriate.
Top University
Yuhang Li · Yuchen Li
Beihang University · East China Normal University
Research Digest··2 min read
In one line
RD-Forget separates storage from use so agents retain all observations but only use query-relevant evidence for each answer.
Before you read this
An idea this paper builds onRead it as
Li and Li introduce RD-Forget, a training-free memory framework that separates long-term storage from the evidence used to answer each query.
The authors retain all observations in a source archive, then use a frozen language model to build a smaller, query-conditioned memory view at answer time.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.