acceptodds
Withdrawn from ICLR 2027

Can your Attractors Adapt? : Persistent Read and Write in Associative Memory

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Associative memory networks recover stored information by evolving incomplete cues towards learned attractors. Most existing formulations assue a fixed memory bank and therefore leave a central problem unsolved : How can memories be added without unpredictable damaging the landscape that supports future recall? Persistence has the properly of changing the dynamics of all future memories during ingestion. For this reason, we introduce **Attractor Lifecycle Control ( ALC )**, a framework for persistent read and write operations over a versioned energy landscape and reading as controlled navigation through it. During recall, a transient query lens reshapes the geometry seen by the dynamics without altering the persistent base landscape. During writing, candidate updates are evaluated for their effects on memory separation, attractor stability and perviously stored content before they can modify the system. To produce effective read and write proposals, ALC uses a recurrent **Proposal Flow Model ( PFM )** trained through continuous-conditional flow matching. PFM iteratively refines schedules and candidate write transformations using the current landscape, operation constraints and its previous predictions. Its outputs remain proposals : the dynamical system executes them, observes their consequences, and accepts only outcomes that satisfy the corresponding geometric and stability checks. ALC achieves **82.1% Recall@1** on LongMemEval-S and reaches an exactly certified stored core on **91.8%** of queries. On controlled synthetic tasks derived from approximately 770,000 arXiv abstracts, ReadPFM improves certified retrieval success by **25%**, while WritePFM enables exact writes for **28.1%** of initially uncertifiable cases. Cost-fitted physical descent reduces NFE by **72.2%** during inference without changing terminal outcomes.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
associative memory, persistent memory, energy-based models, flow matching, nonlinear control, verified learning, dynamical systems, continual learning

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_QuZmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a framework to solve the continual-write problem in associative memories, where the goal is to update an associative memory to store incoming data such that the new data point and the previously written data points remain retrievable. This paper addresses the problem under the growing-capacity assumption, where the framework's on-disk storage increases with every write. The proposed framework, dubbed Attractor Lifecycle Control (ALC), intuitively makes sense - it creates an embedding of the data point on the energy landscape and only accepts the proposal based on certain conditions that enable successful recall of the new and existing points. Otherwise, it uses a model to propose possible updates to the initial location of the new data point. Reads from the AM for a given query start with an immutable version of the energy landscape and a compute budget. The query text is embedded into the same space as the keys. A frozen scorer compares the query against all K keys and produces: a score per memory (essentially a softmax-separated dot product, identical to attention with temperature), one hypothesis, i_q, the memory it believes is wanted, or a refusal, if it cannot name a unique candidate. The state is updated via gradient descent on a temporarily biased energy that favours the scorer’s hypothesis, and the bias is removed in stages so that the final state is judged on the unmodified landscape. Every iteration is checked for numerical sanity, and the budget is reduced after each iteration. The model is then shown the actual endpoint of the current attractor and re-plans the remaining stages. If the gradient is still above tolerance in the final stage and budget expires, the read is marked as a failure. The read succeeds only if the ball it ended in is the one the scorer named. Anything else (a different ball, no ball, a numerical failure, budget exhausted) is returned as a failure.

Critical strengths
  • The paper targets an open and interesting problem. Writing continually to an associative memory without unpredictably damaging earlier recall is unsolved, and there is little theory for it. Most continual memories, including BayesPCN, GDN-2 and SSM-style variants, are defined by an update equation chosen on intuitive grounds, and the effect of a write on the energy landscape is observed afterwards. This paper asks the question directly: what does a write do to the landscape, and under what conditions do earlier attractors survive?

  • It gives a concrete energy landscape-level account of a write. Table 2 identifies which fixed-bank guarantees break when the bank changes and why. Appendix A.7 shows that the energy derivative acquires sign-indefinite terms under a moving bank, so dissipation cannot be claimed through a write. C0–C4 then state explicit, checkable conditions under which every existing attractor provably persists after an insertion. I am not aware of a comparable statement in the continual associative-memory literature.

  • It makes the capacity constraint explicit. For fixed design constants, the separation condition implies a finite packing bound (A.16), so the framework states where a bounded landscape runs out of room. I would like to see more experimental results on capacity bounds and falloff beyond capacity, though, and ideally see a graceful decay rather than immediate collapse.

Critical weaknesses
  • The approach is never compared with the heuristic memories it would replace in the field. The paper’s claim is that a landscape-level theory of writes is preferable to intuitive update rules. That requires a comparison with those rules on a shared continual-write task. None is provided: no BayesPCN (Yoo & Wood, 2022, not cited), GDN/2 or SSM variants, none of the cited continual memories (Bartunov et al.; Alonso & Krichmar), and no plain modern Hopfield update with simple concatenation-based writes.

  • (See the last comment in strengths) The framework says when a write is safe, but not how to keep writing once the landscape is full. Parametric continual memories such as BayesPCN hold memories in a fixed set of weights and must absorb interference. ALC stores every key explicitly in a matrix that grows with K, so storage is not fixed-capacity in the usual sense; the bounded resource is room in the landscape. When that room runs out, the only response is rejection. There is no mechanism for forgetting, consolidating, overwriting or re-placing existing memories to admit a new one. Write repair moves only the incoming key, within a small bound. Preservation of earlier content is therefore close to guaranteed by construction, while the cost of limited capacity is shifted onto new writes. Rejected writes are a form of forgetting that the paper does not count as such. The paper thus addresses only a part of the problem it motivates: a criterion for safe insertion, without a policy for continual operation at capacity. The second half is where heuristic methods differ most and where a landscape-level theory would be most valuable.

  • Capacity and scaling are not examined empirically. All banks are smaller than the embedding dimension (n = 384). This makes perfect recall near-trivial, especially in cases of orthogonal memories. This is a common problem with all fixed-capacity memories that I am aware of: they can easily store and recall up to the point where the state/memory matrix achieves full rank, then drop off hard in either new writes, older memory retention, or both. To me, it seems that ALC would fall into the first category, bringing me back to my first weakness: while interesting, what does the theory of ALC actually provide in practice over other memory models?

  • Writing: Section 3 relies on undefined objects (predictaes, scorer etc). S, Q, R and C are each reused for unrelated quantities and are not defined in the section itself, making it impossible to understand the paper without the appendix. The appendix is not to be used for core paper contributions.

  • At the high level, my main concern can be summarized as the theoretical side being interesting, but the applicability in practice and empirical advantages over other existing frameworks being unstated intuitively and completely absent experimentally. I am happy to discuss this further and increase my score if the authors can actually show practical benefits of applying the ALC framework when compared to existing continual AMs, especially when the number of stored patterns exceeds the dimensionality (384 in the paper's experiments).

Questions for authors
  • What does physical descent provide beyond top-1 lookup (for example, by simple similarity-based metrics)?
  • Can the framework remove or relocate an existing memory to admit a new one, with the certificate re-established? What would the analogue of BayesPCN’s forget operation look like?
  • How is the frozen scorer defined, and why does it differ from raw MiniLM cosine?
Reviewer_FGevOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Attractor Lifecycle Control (ALC), a framework for persistent read and write operations in energy-based associative memory. It combines query-conditioned retrieval, verified memory updates, and recurrent flow-matching models for generating read and write proposals. The authors provide theoretical analyses of local stability and energy dissipation, and evaluate the framework on LongMemEval-S and controlled associative memory tasks.

Critical strengths

Meaningful problem formulation. The paper addresses an important challenge in associative memory: maintaining retrieval stability while continuously incorporating new memories.

Principled framework. The separation between learned proposals, dynamical execution, and verification provides a systematic approach to controlling memory updates, supported by theoretical analysis and empirical evaluation.

Critical weaknesses

Limited evidence for the necessity of the proposed learning components. The framework combines established techniques, including flow matching, nonlinear control, and stability verification. Although their integration is meaningful, the experiments do not sufficiently establish whether the recurrent flow-matching models provide substantial advantages over simpler optimization or proposal-generation alternatives under comparable computational budgets.

Limited validation of generality. The theoretical guarantees concern local properties under specified assumptions, while the learned read and write proposals are primarily evaluated in controlled settings. The relatively small evaluation sets for these components limit the evidence for their robustness and generalization across different memory configurations.

Questions for authors

How does PFM compare with simpler optimization-based or learned proposal methods under matched computational budgets? What evidence demonstrates that recurrent flow matching is necessary for the reported improvements?

How well do the learned read and write proposals generalize to memory configurations, sizes, or update sequences beyond those used in the controlled evaluations?

Reviewer_dJ9pOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposed "persistent associative memory" which is a life cycle control over the energy landscape with different versions. Reading from the memory system is guiding the query in the fixed energy landscape, while writing to the system must pass a certification test which ensure that the new writes create meaningful separations. Furthermore, a flow-matching based model is introduced to suggest read schedule and write admission.

Critical strengths
  1. Verifying memory writing is ensures the stability of the energy landscape and is helpful to sustainable memory retrieval. The idea of checking memory writes keeps the energy landscape in good shape, as adding new memory might changes earlier memory attractors (stored patterns in the landscape), and adding only verified writes keep earlier memories persistent.
  2. The C0-C4 conditions seems solid for memory write certificate.
Critical weaknesses
  1. The paper is hard for me to read. New notations are jumping out everywhere without definitions in the main text (things before appendix). For example, in Section 3.1, RS\mathbf R_S is used without definition, and what is SS in the subscript? Then, a bunch of new symbols appears: Vv=(Kv,Xv,Yv,ρv,Iv,Pv;D,ψ,hv)\mathcal V_v = (K_v, \mathbf X_v, \mathbf Y_v, \boldsymbol \rho_v, \mathcal I_v, \mathbf P_v ; \mathbf D, \psi, h_v) (Eq. 2, in line 170-171), and it says "ρv\boldsymbol \rho_v, Iv\mathcal I_v, Pv\mathbf P_v, and hvh_v bind radii, ordered identifiers, payload lineage, and version content;" (line 172-173). And it explais nothing, what is "radii, ordered indentifiers, payload lineage, version content"? There is no clue for these, and no context. Then, Eq. 3 is also confusing, what is Inv here mean? And the "C0-C4 condition" is showing up every where in the main text but it is defined in Appendix A.9, Eq. 54. Section 3.2 is also confusing to read, as well as Fig. 2 and 3. The "control step" and "Replan" in Read and "Conflict" and "Repair" in Write need to explain in a more detailed way, what leads to such conditions? and how the model deal with it? Current versions of Fig. 2 need people to guess.
  2. The experiment results are difficult to understand. It seems like this work is more in a theoretical way, and it can be improved by providing practical improvement in persistent memory tasks.
  3. The related work should discuss more recent work on managing and manipulating energy landscape, which should include (at least) the line of recent advancement in modern Hopfield network: [1]'s generalized sparse Hopfield network (using sparsemax for separation), and [2]'s adaptive Hopfield network (finding a adaptive way for similarity function). Please distinguish retrieval sparsity in [1] or adaptive similarity [2] from ALC’s query-conditioned guidance and certified persistent writes. This distinction is particularly important because convergence to a certified attractor does not establish correct retrieval (the retrieved pattern is the intended one).
Questions for authors

See the weakness part.

References

[1] Hu, J.Y.C., Yang, D., Wu, D., Xu, C., Chen, B.Y. and Liu, H., 2023. On sparse modern hopfield model. Advances in neural information processing systems, 36, pp.27594-27608.

[2] Wang, S., Pan, Y., Shen, Z., Zhang, M., Wang, H. and Li, G., 2026, April. Adaptive hopfield network: Rethinking similarities in associative memory. In International Conference on Learning Representations (Vol. 2026, pp. 116942-116974).

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

Related papers

Discussion (0)

Sign in to comment.