OpenNovelty.org

Evoking User Memory: Personalizing LLM via Recollection-Familiarity Adaptive Retrieval

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
Large Language ModelMemory RetrievalRecollection-Familiarity Dual ProcessPersonalization
Abstract:

Personalized large language models (LLMs) rely on memory retrieval to incorporate user-specific histories, preferences, and contexts. Existing approaches either overload the LLM by feeding all the user's past memory into the prompt, which is costly and unscalable, or simplify retrieval into a one-shot similarity search, which captures only surface matches. Cognitive science, however, shows that human memory operates through a dual process: Familiarity, offering fast but coarse recognition, and Recollection, enabling deliberate, chain-like reconstruction for deeply recovering episodic content. Current systems lack both the ability to perform recollection retrieval and mechanisms to adaptively switch between the dual retrieval paths, leading to either insufficient recall or the inclusion of noise. To address this, we propose RF-Mem (Recollection–Familiarity Memory Retrieval), a familiarity uncertainty-guided dual-path memory retriever. RF-Mem measures the familiarity signal through the mean score and entropy. High familiarity leads to the direct top-KK Familiarity retrieval path, while low familiarity activates the Recollection path. In the Recollection path, the system clusters candidate memories and applies α\alpha-mix with the query to iteratively expand evidence in embedding space, simulating deliberate contextual reconstruction. This design embeds human-like dual-process recognition into the retriever, avoiding full-context overhead and enabling scalable, adaptive personalization. Experiments across three benchmarks and corpus scales demonstrate that RF-Mem consistently outperforms both one-shot retrieval and full-context reasoning under fixed budget and latency constraints. Our code can be found in the Supplementary Materials.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper proposes RF-Mem, a dual-process memory retrieval framework that distinguishes between familiarity-based and recollection-based retrieval paths for personalized LLMs. It resides in the 'Dual-Process and Adaptive Retrieval' leaf under 'Personalized Retrieval and Generation Methods'. Notably, this leaf contains only one paper in the taxonomy—the original work itself—indicating a sparse research direction. Among the 50 papers surveyed across the field, no other work occupies this specific niche, suggesting that cognitively inspired dual-process retrieval for personalized LLM memory is an emerging area with limited prior exploration.

The taxonomy reveals that neighboring leaves pursue related but distinct strategies. 'Graph-Based Personalized Retrieval' emphasizes structural relationships in user data, while 'Personalized Query Expansion and Optimization' focuses on reinforcement learning and search refinement. The broader 'Memory Architecture and Management Systems' branch includes hierarchical frameworks like MemoryBank and MemGPT, which organize memory storage but do not explicitly model familiarity-recollection dynamics. RF-Mem's cognitive grounding differentiates it from these structural and optimization-centric approaches, positioning it at the intersection of personalization and human-like memory processes.

Among the 20 candidates examined via semantic search and citation expansion, none clearly refute the three core contributions. The dual-process framework itself was assessed against 10 candidates with zero refutable overlaps, as was the familiarity uncertainty-driven selection mechanism. The recollection clustering component was not examined against any candidates. This limited search scope—20 papers rather than an exhaustive survey—means the analysis captures top semantic matches but may not cover all relevant prior work. The absence of refutations among examined candidates suggests novelty within this search window, though broader literature may contain related ideas.

Based on the available signals, RF-Mem appears to introduce a distinctive approach within a sparsely populated research direction. The taxonomy structure and contribution-level statistics indicate that, among the examined candidates, no prior work directly anticipates the familiarity-recollection dual-path mechanism. However, the limited search scope (20 candidates) and the single-paper leaf status warrant caution: the novelty assessment reflects top-K semantic proximity, not exhaustive field coverage. Further investigation into cognitive science-inspired retrieval and adaptive memory systems may reveal additional context.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
20
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: Personalized memory retrieval for large language models. The field has organized itself into several major branches that reflect different facets of how memory and personalization can be integrated into LLM systems. Memory Architecture and Management Systems focuses on the underlying infrastructure—how to store, organize, and efficiently access memory modules, with works like MemoryBank[5] and MemGPT[32] exploring hierarchical and operating-system-inspired designs. Personalized Retrieval and Generation Methods emphasizes adaptive strategies that tailor retrieval to individual user contexts, including dual-process mechanisms and cognitive-inspired approaches such as Cognitive Memory[3] and Cognitive Personalized Search[7]. Application-Specific Personalization targets domain-driven use cases like conversational recommendation (MemoCRS[9]) and educational retrieval (Exercise Retrieval[15]), while General Retrieval-Augmented Generation and Knowledge Integration addresses broader RAG frameworks and knowledge grounding. Hardware and Inference Optimization tackles the computational challenges of serving memory-augmented models at scale, with innovations like PagedAttention[4] and vAttention[12]. Finally, Surveys, Benchmarks, and Cross-Cutting Studies provide evaluative perspectives and comparative analyses across these dimensions. Within Personalized Retrieval and Generation Methods, a particularly active line of work explores dual-process and adaptive retrieval strategies that mirror human cognitive mechanisms. Recollection Familiarity Retrieval[0] sits squarely in this cluster, drawing on dual-process theories to distinguish between detailed recollection and gist-based familiarity when retrieving personalized memories. This contrasts with more graph-centric approaches like Personalized Graph Retrieval[1], which emphasizes structural relationships in user data, and with cognitively inspired frameworks such as Cognitive Memory[3], which integrates episodic and semantic memory layers. A key trade-off across these methods is the balance between retrieval precision and computational overhead: while dual-process models aim for nuanced, context-sensitive recall, simpler architectures may sacrifice granularity for speed. Open questions remain around how to dynamically adjust retrieval strategies based on query complexity and user history, and how to evaluate the subjective quality of personalized outputs in diverse application settings.

Claimed Contributions

RF-Mem dual-process memory retrieval framework

The authors introduce RF-Mem, a retrieval framework that adaptively switches between fast Familiarity retrieval and deliberate Recollection retrieval based on familiarity uncertainty signals. This design embeds human-like dual-process recognition into memory retrieval for personalized LLMs.

10 retrieved papers
Familiarity uncertainty-driven adaptive selection mechanism

The authors develop a selection mechanism that measures familiarity through mean score and entropy, enabling the system to adaptively choose between one-shot Familiarity retrieval and stepwise Recollection retrieval based on confidence levels.

10 retrieved papers
Recollection retrieval via clustering and query-centroid mixing

The authors propose a stepwise Recollection retrieval method that clusters candidate memories using KMeans, mixes cluster centroids with queries through alpha-mixing, and iteratively expands evidence chains to simulate deliberate contextual reconstruction.

0 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

RF-Mem dual-process memory retrieval framework

The authors introduce RF-Mem, a retrieval framework that adaptively switches between fast Familiarity retrieval and deliberate Recollection retrieval based on familiarity uncertainty signals. This design embeds human-like dual-process recognition into memory retrieval for personalized LLMs.

Contribution

Familiarity uncertainty-driven adaptive selection mechanism

The authors develop a selection mechanism that measures familiarity through mean score and entropy, enabling the system to adaptively choose between one-shot Familiarity retrieval and stepwise Recollection retrieval based on confidence levels.

Contribution

Recollection retrieval via clustering and query-centroid mixing

The authors propose a stepwise Recollection retrieval method that clusters candidate memories using KMeans, mixes cluster centroids with queries through alpha-mixing, and iteratively expands evidence chains to simulate deliberate contextual reconstruction.

Table of Contents

Evoking User Memory: Personalizing LLM via Recollection-Familiarity Adaptive Retrieval | Novelty Validation