← All AI/ML open problems

Memory & capacity · RMT & weight spectra / Open question

How Many Memories Can Coexist?

Associative recall must recover the right memory despite noise and interference. Correlated real-world memories make this harder than simply counting possible stable states.

OpenEditorially scopedFramework needs definitionCompiled 2026-11-19

Seed catalog · Showing the bundled record.

Problem statement

Generate binary memories by choosing one of K random prototype vectors in dimension d and independently flipping each coordinate with probability q below one half; corrupt retrieval cues by a further flip probability eta. Insert memories sequentially into a modern Hopfield network with a stored pattern matrix and softmax retrieval at inverse temperature beta, all stored entries having p-bit precision. Determine the sharp load/correlation/noise frontier for recall of a uniformly selected stored memory, with basin size, iteration count and collisions or insertions accounted for.

Assumptions and scope

  • Fix the correlation/generative model, cue corruption law, update dynamics, learning rule and definition of successful recall.
  • Distinguish independently prescribed memories from attractors generated by a compact construction; many attractors do not imply storage of equally many independent facts.
  • Charge all trainable weights, explicit patterns and helper memory. Exponential capacity in state dimension does not imply exponential information per stored bit.

What would settle it?

A sharp load–correlation–noise frontier for a specified learned-memory model, with matching achievable retrieval and interference bounds over an incremental learning process.

Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.

Status and evidence

Modern Hopfield models already have exponential pattern-count and retrieval guarantees under stated conditions. Asymmetric binary constructions also achieve many long noise-robust cycles. The residual target is learning and recalling a prescribed correlated family under fully charged storage, not a first high-capacity attractor construction. This is an editorial, source-motivated research target. The preliminary review through 2026-11-19 does not certify novelty or exclude an unlocated resolution.

Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.

Sources and further reading

  1. Ramsauer et al. — Hopfield Networks Is All You Need ↗partial theory · source checked 2026-11-19

    Primary abstract checked: exponential pattern counts in associative-space dimension, retrieval bounds and attention equivalence in the stated modern Hopfield model. Pattern matrices and their precision still count toward total storage.

  2. Kumar et al. — Beyond Fixed Points: Superpolynomial Capacity of Asymmetric Hopfield Networks ↗partial theory · source checked 2026-11-19

    Primary abstract checked: constructed binary synchronous networks with many long noise-robust limit cycles. Attractor counts for a designed network are not the capacity to learn arbitrary independent memories.

  3. Wong — Paying Attention to Facts: Quantifying the Knowledge Capacity of Attention Layers ↗partial theory · source checked 2026-11-19

    Primary search abstract checked: database/attention tensor-rank bounds and empirical relationships for one-layer factual memorization. Expressivity, learned storage and reliable recall are distinct.