What they did
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 curator extracts relevant evidence, groups facts into semantic slots, links newer values to the same-slot values they replace, and retains relations needed for multi-step reasoning.
A rate-distortion objective—which balances memory size against loss of useful information—guides evidence selection under a fixed memory budget. The shared answering pipeline was evaluated on conversational memory, knowledge updating, fact consolidation, long-context reasoning, and personalization tasks.
Key findings
- Configurations lacking selective forgetting or query conditioning produced the largest score deficits in the reported ablations.
- Replacement links helped prevent superseded values from influencing questions about the current state.
- Intent-aware retrieval restored older evidence for historical questions rather than deleting it permanently.
- Semantic-slot grouping and relation preservation made complementary contributions, including support for multi-hop reasoning.
Why it matters
The work reframes forgetting as control over information use rather than deletion from storage. This distinction could help persistent agents handle changing user preferences and world facts without sacrificing historical recall or contaminating current answers with obsolete alternatives.
Caveats
RD-Forget depends on a frozen language model to identify evidence, semantic slots, replacement relations, and query intent, so curator errors could propagate to the final answer. The abstract reports broad evaluations and relative ablation trends but provides no numerical effect sizes, leaving efficiency, robustness across model families, and performance under ambiguous or contradictory updates as open questions.