TRACE-Memory: Public-Conditioned Retrieval and Utility-Aware Evidence Admission for Personalized Generation

2026-08-08

AI Summary

Existing personalized generation methods often produce ineffective or even harmful outputs due to retrieved user histories that deviate from relevant preference dimensions, redundantly reiterate public information, or lack sufficient evidential support. To address this, this work proposes TRACE-Memory, a two-stage framework that first constructs a coverage-oriented candidate memory pool to complement missing information in the public context, and then selectively incorporates a verifiable subset of evidence—or none at all—based on response-level incremental utility, thereby enabling utility-driven personalization. The approach innovatively conditions memory usage on its incremental contribution over the public context, introducing structured evidence units and a utility-aware admission mechanism. Experiments across 4,500 tasks on Goodreads, Amazon Reviews, and Reddit demonstrate that TRACE-Memory significantly outperforms multiple retrieval baselines and adaptively modulates memory utilization according to contextual sufficiency, matching the memory-based performance of state-of-the-art large models when enhanced with local model capabilities.

Abstract

Personalized generation systems retrieve user history by request--memory relevance and inject it into the model context. Yet relevant history may concern the wrong preference aspect, duplicate public information, or provide insufficient support. We argue that personal memory should be used only when it adds utility beyond a public-only response. We propose TRACE-Memory, a two-stage framework for selective personalization. Stage 1 queries for user-specific information missing from the request and public context, then retrieves a coverage-oriented candidate pool. Stage 2 admits a compact subset of source-traceable evidence units, or the empty set, according to response-level incremental utility. We progressively train the query-generation and evidence-admission policies through structured SFT initialization, reduced-space stage-wise GRPO warm-up, and nested multi-sample Joint GRPO. Across 4,500 Controlled and Natural tasks from Goodreads, Amazon Reviews, and Reddit, TRACE-Memory consistently outperforms random and lexical memory use, improves over semantic retrieval, remains competitive with frontier-LLM memory pipelines as local generator capacity increases, and conditions evidence admission on public-context sufficiency, supporting selective rather than default personalization.

Problem

• personalized generation
• memory retrieval
• utility-aware
• evidence admission
• public-conditioned

Innovation

• selective personalization
• utility-aware evidence admission
• public-conditioned retrieval
• incremental utility
• source-traceable evidence