Technical Summary: Memory Retrieval for Changing Preferences

Problem Statement

Long-context dialogue systems face a critical challenge in personalized AI: deciding when to access external memory and which parts of a user's interaction history are relevant. Existing approaches typically rely on heuristic retrieval signals or "always-on" memory usage, often failing to account for the changing and potentially inconsistent nature of user preferences.

The standard practice of using semantic embedding similarity (selecting turns nearest to the query in embedding space) is identified as a structural limitation. Semantic similarity measures topical co-occurrence but fails to distinguish between:

1. Evidential utility: Whether a memory turn provides incremental evidence to resolve the current query.
2. Topical relatedness: Whether a turn is merely about the same subject.

This mismatch leads to retrieval failures where systems surface context that is topically related but evidentially uninformative (e.g., retrieving an old preference that has since been superseded) or miss sparse, temporally distant evidence crucial for the correct answer. Furthermore, existing systems often treat memory access (gating) and memory selection as independent decisions, neither grounded in a direct estimate of memory utility for the specific query.

Methodology

The authors propose a unified framework for memory access and selection based on changing preferences, operationalized through a Bayesian lens. The core idea is to quantify the utility of each memory turn as the improvement in the model's likelihood of the reference response when that turn is included.

1. Theoretical Foundation: Bayes Factor as Utility

The framework defines memory retrieval as identifying historical turns that provide evidence about a user's latent preference state.

• Answer-Conditioned Saliency ($Sal_a$): Defined as the log Bayes factor, measuring how much observing a turn $c_i$ updates the model's belief about the answer $a$ given query $q$:
  $$Sal_a(c_i | q, a) = \log p(a | q, c_i) - \log p(a | q)$$
  This serves as an "oracle" signal indicating genuine memory evidence.
• Query-Conditioned Proxy ($Sal_q$): Since the gold answer $a$ is unavailable at inference time, the authors derive a tractable proxy by replacing $a$ with $q$:
  $$Sal_q(c_i; q) = \log p(q | c_i) - \log p(q)$$
  Theoretically, under the condition of query-conditional sufficiency (where the turn encodes the necessary user-state information), $Sal_q$ preserves the ranking of $Sal_a$.

2. Four-Stage Pipeline

The method implements a four-stage process to unify access and selection:

• Stage 1 (Oracle Focus): Defines an ideal set of turns $S_\tau$ where $Sal_a > \tau$.
• Stage 2 (Proxy Construction): Uses $Sal_q$ as a surrogate for $Sal_a$ to pre-filter the conversation history.
• Stage 3 (Threshold-Gated Memory): Converts the continuous saliency signal into a discrete decision. A threshold $\tau$ determines:
  • Gating ($g_\tau$): Whether to trigger memory access at all.
  • Selection ($S_\tau$): Which turns to retain.
• Stage 4 (Learning the Policy): A Supervised Fine-Tuning (SFT) pipeline trains a single model with two heads:
  • Controller Head: Takes the query and candidate evidence, outputs a structured action (trigger gate + focus indices) based on the silver labels derived from $Sal_a$.
  • Reader Head: Answers the query (MCQ) using the selected evidence.
    Both heads share parameters and are trained jointly on a next-token prediction objective.

3. Inference

At inference, the system:

1. Pre-filters the history using $Sal_q$ to reduce the search space.
2. The Controller (trained on $Sal_a$ labels) predicts whether to access memory and which turns to select, without computing $Sal_a$ directly.
3. The Reader generates the response conditioned on the selected turns (or the query alone if the gate is off).

Key Contributions

1. Principled Retrieval Signal: The paper introduces a Bayes-factor-based signal that captures evidential utility rather than topical relatedness, offering a principled alternative to semantic similarity.
2. Unified Framework: A "Bayes-factor-gated" framework that couples memory access (gating) and turn selection (saliency) under a single signal, addressing the decoupled nature of prior systems.
3. Distinctiveness of Signals: Controlled ablations demonstrate that Bayes-factor-based saliency labels are not interchangeable with embedding-based labels. The gap is most significant on long-context, preference-intensive tasks.
4. Improved Long-Context Personalization: The approach demonstrates consistent improvements across benchmarks, particularly where semantic similarity fails to capture implicit or evolving preferences.

Experimental Results

The method was evaluated on four heterogeneous benchmarks: MemBench-Low/High, PersonaMem, and PrefEval.

• Overall Performance: The proposed method outperformed all retrieval-augmented baselines (including Mem0, A-MEM, MemoryBank, and RMM) and the full-context baseline.
  • On Qwen3.5-9B-Thinking, it achieved an average accuracy of 84.94, surpassing the strongest baseline (MemoryBank at 65.64) by +19.30 points.
  • On Llama-3.1-8B-Instruct, it reached 82.05 vs. 71.65 for the best baseline.
• Task-Specific Gains:
  • The largest improvements were observed on MB-High and PersonaMem, which require implicit inference and reasoning over long-range, changing preferences. For instance, on PersonaMem, the method improved accuracy by +19.97 points over the best baseline.
  • Gains were smaller on PrefEval, where preferences are often explicit and semantic retrieval is already effective.
• Ablation Studies:
  • SFT is Essential: Removing supervised fine-tuning and using the base model with oracle scoring collapsed performance, proving the utility signal must be internalized.
  • Signal Quality: Replacing $Sal_a$ labels with embedding-based pseudo-labels reduced average accuracy by 2.39 points, with a 8.76 point drop on MB-High, confirming that evidential utility is distinct from semantic similarity.
  • Controller Role: The controller provided marginal but consistent benefits, particularly on harder retrieval scenarios (MB-High).

Significance and Claims

The paper claims that effective personalized memory retrieval depends less on maximizing topical coverage and more on identifying evidence with high answer utility. By framing retrieval as utility estimation via Bayes factors, the system learns to:

• Identify salient turns that reflect changing user states.
• Regulate memory usage based on expected utility, avoiding the injection of irrelevant context.
• Operate effectively in long-context regimes where semantic similarity is insufficient.

The authors position this work as a shift from heuristic, similarity-based retrieval to a principled, evidence-based approach that better aligns with the requirements of dynamic personalization. The framework demonstrates that jointly training retrieval and reasoning components allows the system to adapt to compressed evidence representations, achieving superior performance without requiring gold answers at inference time.