Source-linked AI summary
What Should an Agent Forget? Separating What Is Stored from What Is Used
Yuhang Li, Yuchen Li
TL;DR
Persistent agents must retain experience while using only evidence suited to the current question, because superseded facts may be wrong for current-state answers but useful historically. RD-Forget separates storage from use with an archived source history and a query-conditioned memory view that curates, links, and budgets evidence. Across diverse tasks, the largest deficits occur without forgetting or query conditioning, supporting selective control of retained history.
Problem
Persistent agents need stored experience across time, but changing facts and query-specific evidence needs make it difficult to decide what should remain stored versus what should influence an answer.
Method
RD-Forget retains source observations in an archive and constructs a query-conditioned memory view using semantic-slot curation, supersession links, relation preservation, intent-aware retrieval, and budgeted selection.
Results
Across five task suites and four language models, RD-Forget improves answer quality over baselines, while matched ablations show the largest deficits without forgetting and without query conditioning.
Takeaways & Limitations
Retaining history while selectively controlling its use supports both current-state answers and historical questions whose evidence requirements change with the query.
Abstract
from arXiv · showhide
Persistent language agents need stored experience to remain available across time, while each answer requires evidence suited to a particular question. A superseded fact can mislead a current-state answer and still be essential for a historical query. We present RD-Forget, a training-free framework that separates what an agent stores from what it uses. A retained source archive preserves observations, and a query-conditioned memory view controls their influence on the current answer. A frozen language-model curator extracts relevant evidence, groups facts into semantic slots, and preserves the relations needed for multi-hop reasoning. Same-slot replacement links suppress superseded values in current-state contexts, while intent-aware retrieval makes earlier evidence eligible again. A rate-distortion formulation guides construction of the answer-time view within a memory budget. Experiments span conversational memory, knowledge updating, fact consolidation, long-context reasoning, and personalization under a shared answering pipeline. The results associate accurate answers with both query-relevant evidence construction and control over obsolete alternatives. Configurations without forgetting or query conditioning have the largest score deficits, while slot grouping, historical access, and relation preservation contribute complementary functions. Retaining history while selectively controlling its use offers a practical way to accommodate changing facts and future questions.
1 Introduction
Persistent agents must distinguish what remains stored from what influences an answer because changed facts can be obsolete for current questions yet useful historically. RD-Forget makes this distinction explicit through query-conditioned evidence selection, relation-aware curation, and budgeted answer contexts.
- Changed facts can mislead current-state answers while remaining useful for historical questions, so storage and answer-time use should be separated.
- RD-Forget retains source observations in an archive while a query-dependent memory view controls their influence on the next answer.
- Semantic slots and same-slot replacement suppress superseded values for current-state questions while preserving complementary relations needed for reasoning.
- A rate–distortion trade-off constructs the answer-time view by selecting evidence that preserves answer utility within a representation budget.
- Across five task suites and four language models, fact-consolidation gains over the better baseline range from 11.00 to 26.00 percentage points.
2 Related Work
Prior memory systems retain, organize, and retrieve experience, while RD-Forget focuses on constructing query-appropriate evidence from retained history. Its scope-aware forgetting addresses both obsolete values and the risk of discarding relations needed later.
- RD-Forget conditions forgetting on the current question and the semantic scope of a replacement rather than treating forgetting as uniform memory decay.
- An old fact can leave the current answer context while remaining accessible for historical use, connecting forgetting to changing evidence requirements across queries.
- Existing systems accumulate observations, manage memory tiers, construct linked notes, or reuse task experience; RD-Forget instead focuses on query-conditioned evidence construction.
- The evaluated task suites expose complementary risks: obsolete values can corrupt current answers, while discarded history or relations can make later questions unanswerable.
3 Separating Storage from Use
RD-Forget derives a query-specific memory view from stored history by extracting relevant evidence, linking replacements, determining eligibility, and packing selected entries under a budget. The design preserves relation chains and supports historical or multi-state questions when intent requires them.
- 3.1 Stored history and used evidence: Stored source history remains available while the answer model receives a derived memory view constructed through evidence extraction, replacement relationships, and eligibility decisions.
- 3.1 Stored history and used evidence: Rate–distortion selection treats omitted evidence as costly when it reduces answer utility, with the query determining which omissions matter.
- 3.2 Evidence curation: The curator extracts atomic facts with semantic slots, making replacement specific to a relation such as employment rather than an unrelated employer-location fact.
- 3.2 Evidence curation: For multi-hop queries, curation retains each relation needed to reach the answer while suppressing a competing previous value.
- 3.2 Evidence curation: Temporal interpretation distinguishes replacements from related observations that can coexist, supporting current-state, historical, and evolution questions.
- 3.4 Reusing stored history: Retrieval eligibility includes active, deprecated, and superseded entries, with historical or revision-chain intent enabling non-active evidence to re-enter selection.
- 3.5 Packing the answer-time view: The pipeline ranks eligible entries, selects an ordered subset, and greedily packs entries whose costs fit the remaining memory budget.
- 3.6 End-to-end procedure: The procedure combines curation, active materialization, supersession linking, eligibility filtering, ranking, selection, and budgeted answer generation.
4 Experimental Setup
The experiments evaluate RD-Forget across conversational memory, knowledge updates, fact consolidation, long-context reasoning, and personalization using multiple models and fixed comparison protocols. Ablations isolate forgetting, rescue, slot grouping, relation closure, and query conditioning under shared task sets.
- Main evaluation suites: The main evaluation uses AMB-Text, LME-KU, and MAB-FC, covering conversational memory, knowledge updates, and fact consolidation.
- Main evaluation suites: The three main suites comprise 264 tasks per method, while query-intent evaluation adds BEAM and PersonaMem tasks.
- Models and protocol: Four language models use the same model family and client configuration for curation, selection, answering, and model-based evaluation within each run.
- Scoring: AMB-Text and LME-KU use mean binary semantic judgments, whereas MAB-FC uses rule accuracy from normalized substring matching.
- Ablations: The matched Full batch and five variants use identical 264 tasks to compare forgetting, rescue, slot grouping, relation closure, and query conditioning.
5 Results and Analysis
RD-Forget achieves the strongest reported accuracy across models and benchmarks, while ablations show that forgetting and query conditioning are especially important. Slot grouping, historical retrieval, and relation preservation provide additional gains.
- RD-Forget leads all twelve model–benchmark combinations in Table 1, with gains over the stronger baseline of 1.16–19.77 points on AMB-Text and 6.41–25.64 points on LME-KU.
- Fact consolidation gains over the better-performing baseline range from 11.00 to 26.00 points across all four backbones, with Luna showing the largest margin.
- RD-Forget has the highest equally weighted three-suite mean for every model: 66.59% for Qwen, 86.71% for Luna, 77.13% for MiniMax, and 72.83% for Kimi.
- Removing forgetting produces the largest matched-Luna deficits: 23.26 points on AMB-Text, 33.33 on LME-KU, and 23.00 on MAB-FC.
- Removing query conditioning produces the second-largest component-ablation deficits, despite a larger curation allowance and a query-aware downstream selector.
- Closure, rescue, and slot ablations also reduce scores, indicating complementary value from multi-hop relation preservation, historical eligibility, and same-slot grouping.
6 Stored History across Query Intents
The query-intent evaluation extends RD-Forget to evolving knowledge and preferences over long histories. RD-Forget leads every reported model–benchmark combination under the stated scoring protocols.
- The query-intent evaluation covers BEAM knowledge updates and reasoning categories plus PersonaMem preference following, using four backbones.
- RD-Forget leads all eight model–benchmark combinations in the query-intent evaluation.
- On BEAM, the advantage over the better-performing baseline ranges from 8.47 to 15.67 points across the four backbones.
- On PersonaMem, the advantage over the better-performing baseline ranges from 10.88 to 20.72 points across the four backbones.
- The query-intent results extend the comparison to long histories and evolving preferences.
7 Discussion
RD-Forget treats forgetting as selective control over retained evidence rather than deletion of history. The discussion links performance to preserving relation paths while resolving competing values within semantic slots.
- Separating storage from use lets superseded facts leave current-state contexts while remaining available for historical queries.
- Matched Luna ablations rank forgetting as the most consequential component and query conditioning as the next, while closure, rescue, and slot effects are smaller.
- Fact consolidation requires an evidence view that preserves a usable relation path while resolving revisions at each hop.
- The retained archive supplies observations, while curation determines which relations and versions enter the answer context.
- Slot granularity separates complementary facts from competing values of one relation through curator grouping and materializer checks.
8 Conclusion
RD-Forget separates retained history from the evidence used for an answer, allowing current-state responses to exclude obsolete alternatives while preserving historical observations. Figure 3 shows that removing forgetting, followed by removing query conditioning, produces the largest deficits across all three benchmarks.
- RD-Forget retains source history while its memory view governs which evidence influences each answer.
- −FORGET has the largest deficit relative to matched Full on every benchmark, followed by −QC.
- The separation accommodates both current-state evidence and historical observations whose usefulness changes with the question.