acceptodds
Under review as a conference paper at ICLR 2027

PRAM: Partial-Cue Retrieval with Associative Memory over Frozen LLM Representations

Abstract

Long-term language agents need to recover related experiences from partial cues, compose them across relations, and selectively retain information as new experiences arrive. We investigate whether fixed-size linear associative maps over frozen language-model representations can support these functions. Our framework, Partial-cue Retrieval with Associative Memory (PRAM), binds whitened entity representations with relation and time codes and organizes facts into complementary maps for target, set, and temporal access. Successive association reads construct a query-specific tree. A learned bridge converts its target representations into soft tokens, which are interleaved with textual relation structure and supplied to a frozen language model for answering. We further study a dynamic extension in which fast acquisition, decay, and error-driven consolidation regulate the persistence of associations. Through controlled retrieval, neural readout, conversational question answering, and streaming-memory experiments, we characterize the roles of representation geometry, structured access, and selective retention. The framework connects graph-like memory operations with a continuous language interface and provides a basis for studying both how associations are recalled and how they are maintained.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM, Memory, KG

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_f49FOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Long-term language agents need to recover related experiences from partial cues, compose them across relations, and selectively retain information as new experiences arrive. In this work, PRAM stores structured facts in associative matrices and retrieves relational trees for a frozen LLM through a learned soft-token bridge. Separate streaming experiments study selective retention.

Critical strengths
  1. Representing memory through associative matrices with encoded relations and time is an interesting idea, offering a compact representation of structured associations.
Critical weaknesses
  1. Supported query operations need clarification. MetaQA explicitly uses bidirectional storage, but the general construction does not explain whether incoming facts automatically receive reverse associations and inverse relation codes. In Figure 1, storing Leo as Mia’s mentor supports retrieving Leo from Mia; answering “Who has Leo as a mentor?” requires reverse access. More complex queries also remain unclear: A -> B -> X <- C requires finding a shared target satisfying both paths, whereas the described tree expansion unions retrieved neighborhoods. How are inverse relations and such intersection constraints handled during storage and retrieval?

  2. Evaluation does not establish an advantage over existing memory methods. The paper lacks direct comparisons with established memory systems discussed in its related work. Its LoCoMo performance is also weak: PRAM achieves 38.0% accuracy, substantially below the reported dense-RAG baseline’s 58.4%. Furthermore, category 3 is excluded without explanation, narrowing benchmark coverage. Together, these limitations leave the framework’s practical value insufficiently supported.

  3. Fact construction is insufficiently described. The paper evaluates memory over an LLM-extracted fact collection without specifying the extraction procedure or assessing what conversational information it preserves. Consequently, preservation of emotions, motivations, outcomes, and causal context remains unverified.

Questions for authors

see weakness

Reviewer_aYCUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposed a PRAM framework that uses multiplicative binding and complementary linear associative maps over frozen language-model representations to support partial-cue retrieval and construct query-specific association trees. A learned bridge converts retrieved target representations into soft tokens, interleaved with textual relation and branch structure, for answering with a frozen language model. The paper also studies a dynamic extension in which fast acquisition, decay, and error-driven consolidation regulate the retention of associations under continuing writes.

Critical strengths
  1. PRAM combines multiplicative binding with complementary linear maps to support partial-cue, set, temporal, and multi-hop access, providing a coherent approach to graph-like retrieval over frozen language-model representations.

  2. The learned bridge converts target representations into soft tokens and combines them with textual relation and branch structure, enabling a frozen language model to read retrieved association trees.

  3. The controlled retrieval, neural readout, and streaming experiments examine complementary aspects of associative memory, including how decay and fast–slow consolidation support selective forgetting and reinforcement.

Critical weaknesses
  1. The independent effectiveness of error-driven consolidation is not evaluated.

  2. The dynamic-retention comparisons do not match total-state budgets. The dual-layer method uses more storage state than the single-layer baselines, so comparisons at equal total-state budgets are needed to distinguish gains from the write rule from those due to additional capacity.

  3. The main-text evaluation omits the LoCoMo comparison, where PRAM achieves 0.380 accuracy versus 0.584 for dense RAG@40. This result should be included and discussed in the main text to clarify the limitations of structured access in realistic conversational settings.

  4. The evaluation lacks direct comparisons with recent memory systems discussed in the paper, such as Mem0, MemTree, and Zep, or an appropriate learned neural-memory baseline.

  5. The Introduction would benefit from clearer organization. It should focus on specific limitations of existing methods and clarify what associative maps, structured readout, and dynamic consolidation each address, and why they should be combined.

Questions for authors

In Table 2, the four results under “After associative retrieval (Synthetic)” are placed in the PopQA column, whereas the caption, subsection heading, and Section 4.3 describe them as Synthetic results. Could the authors clarify which dataset these results correspond to and correct the column alignment or descriptions accordingly?

Reviewer_YQowOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRAM, a memory framework for LLM agents. The key idea is to store factual knowledge in fixed-size linear associative maps over frozen LLM representations, rather than as text. Entity representations are whitened and then bound with relation and time codes. Retrieved facts are assembled into an association tree, which is converted into soft tokens so that a frozen LLM can answer from it. The paper also introduces fast–slow consolidation so that repeated memories are retained more strongly. Experiments show that binding and whitening are critical for retrieval. PRAM is comparable to symbolic retrieval at shallow hops, but on conversational QA it falls short of dense RAG.

Critical strengths
  1. The problem setting is interesting. The paper asks whether a distributed associative memory can provide the kind of graph-like access that symbolic graph memory offers. I also found it interesting to see classical associative memory and complementary learning systems brought together in one framework.

  2. The analysis is well controlled. The shared-routing comparisons, map ablations, and tree-width experiments each isolate a separate component, and the paper is honest about the limitations of the proposed method.

Critical weaknesses
  1. The benefit of the proposed method is unclear compared to text or symbolic memory. For LoCoMo, PRAM is worse than dense RAG, and comparable or falling behind (for deep hops) with symbolic traversal. The benefit of storage saving is similar to growing KV. The method is also relying on a parse and a candidate entity codebook, which are virtually a symbolic table.

  2. The novelty of the components are limited. Binding, ridge-aligned associative map, fast weight, fast-slow consolidation, and query-based bridge are all well-known techniques. The main contribution is how they are combined but it is rather difficult to follow this as the writing is quite dense.

Questions for authors
  1. Why does PRAM underperform dense RAG in LoCoMo? Given this, why we should use the associative representation? It would be helpful to see a comparison with learned memory systems such as Mem0, Zep, and Larimar under the same storage budget.

  2. How critical is the codebook used for cleanup? If the multi-hop approach doesn't work without the codebook. it may say that the proposed method relies a lot on the entity table rather than the linear map?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.