PaperlayerView exact revision on arXiv

Source-linked AI summary

Retrieval as Traversal: Value-Indexed Memory for Stateful Long-Context Agents

Daniel Strauß, Zhenshan Bing, Alois Knoll

arXiv:2610.40545v1cs.AI

TL;DR

Long-context models repeatedly search retained histories or compress them before future information needs are known, while nearby retrieval steps may use overlapping memories. The paper proposes persistent traversal through an address space organized by retrieval value, and reports a fixed-geometry proof-of-principle experiment that does not validate the learned value-index.

  • Problem

    Long-context models either repeatedly determine which parts of retained histories matter or compress the past before future information needs are known, despite overlap between useful memories at nearby steps.

    Source: §1 · p. 1

  • Method

    The paper separates memory content from address and proposes a cost-regularized decision-theoretic objective that organizes local accessibility by retrieval value.

    Source: §1 · p. 2, §2.1 · p. 2, §5 · p. 6

  • Results

    The proof-of-principle experiment uses fixed geometry and learns only the persistent retrieval mechanism, without experimentally validating the learned value-index.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The paper proposes making valuable memory combinations locally accessible while charging for unnecessary co-retrieval.

    Source: §5 · p. 6

    The experiment uses environment-provided geometry and retrieves only one piece of information, leaving learned geometry and sequential access untested.

    Source: §4 · p. 5

Abstract

from arXiv · show

Long-context models repeatedly solve closely related retrieval problems. Attention or global retrieval can retain large stores, but each forward pass again searches for what matters. Recurrent compression avoids this search only by deciding earlier what information to preserve. Neither mechanism directly exploits that useful memories at neighbouring retrieval steps may substantially overlap. We propose an alternative in which retrieval is a persistent local traversal through an address space. We separate a memory's content from its address. Utility is defined by the information revealed by retrieving the content, while geometry determines only how cheaply that information can be reached. This leads to a cost-regularised indexing objective that preserves information value inside local neighbourhoods while charging for every memory made locally accessible. Its marginal-value form gives both attractive and exclusion pressures. Thus, complementary memories should be co-retrievable when the additional value exceeds retrieval cost, whereas redundant memories should not occupy the same retrieval set solely because they are individually useful. As a proof of principle, we test only the simplest fixed-geometry instance on MiniGrid Memory. A persistent local traversal head reaches $0.95$ mean return after $0.59$M training steps with a five-token context, compared with $0.94$M for Gaussian absolute addressing and more than $2$M for categorical absolute addressing. The experiment does not test learned geometry or repeated retrieval. Instead, it motivates the hypothesis that stateful retrieval may be easier to learn than stateless. To test the full proposed method further experimentation is necessary.

1 Introduction

Long-context retrieval must trade retaining searchable history against compressing information before future needs are known, while successive queries may reuse overlapping memories. The paper proposes persistent traversal through an address space organized by retrieval value, with a decision-theoretic objective and a fixed-geometry proof-of-principle experiment.

1 Introduction

  • Long-context systems either repeatedly search retained histories or compress the past before future information needs are known.Sparse attention and retrieval reduce query costs but typically treat successive queries as fresh matching problems [5] [9] [6].Source: §1 · p. 1
  • Persistent traversal targets overlap between memories useful at neighbouring inference steps by placing context and retrieval heads in an address space.The retrieval head accesses memories within a local distance, so positions are intended to reflect expected retrieval value.Source: §1 · p. 1
  • Addresses should reflect decision value rather than content similarity, because similar memories can be redundant while dissimilar memories can be useful together.The proposal distinguishes memory content from the geometry that determines retrieval accessibility.Source: §1 · p. 2
  • The paper reports a decision-theoretic objective for learning retrieval geometry and a fixed-geometry experiment that learns only the persistent retrieval mechanism.It does not claim experimental validation of the learned value-index.Source: §1 · p. 2
  • The paper frames its approach against memory architectures and retrieval systems, asking whether addresses can encode expected utility and support persistent traversal across related queries.Related work includes Neural Turing Machines, Differentiable Neural Computers, Lie-access Neural Turing Machines [12], Neural Map [8], and retrieval or sparse-attention methods [5] [9] [11] [6].Source: §1 · p. 2

2 Value-indexed memory

The section separates memory value from memory address, then uses local retrieval costs and interactions among memories to shape which contents are co-retrieved. It also describes persistent traversal as a way to reuse retrieval state when useful information persists across steps.

2.1 Content, context, and address

  • Memory value depends on content rather than address, while the current context and retrieval-head position are maintained separately.Geometry determines which contents are cheaply retrievable, not what those contents are worth.Source: §2.1 · p. 2

2.2 A cost-regularised local index

  • A cost-regularised local index balances information omitted from a neighborhood against a cost λ for each accessible memory.The resulting value threshold replaces an arbitrary top-k budget: another memory is accessible when its added information justifies its retrieval cost.Source: §2.2 · p. 3

2.3 Complementarity and redundancy induce opposite pressures

  • Joint retrieval is preferred when the second memory’s conditional marginal value exceeds λ, with positive interaction increasing the contexts where that condition holds.The interaction is decision-theoretic: it concerns changes in marginal value, not raw statistical correlation.Source: §2.3 · p. 3
  • Exact substitutes share interaction profiles and compete for synergistic sets, so address geometry should reflect value and interaction profiles rather than raw content similarity.When one memory makes another redundant, retrieving both incurs the extra retrieval cost; separating alternatives is one possible geometry, not a required solution.Source: §2.3 · pp. 3–4

2.4 Persistent retrieval as traversal

  • Persistent traversal retains recently traversed memories in the working window while the retrieval head continues moving locally.Persistence can reuse retrieval state when useful information persists between consecutive retrievals; the movement rule remains an open design choice.Source: §2.4 · p. 4

3 A highly crude fixed-Φ proof of principle

This proof of principle tests whether persistent relative retrieval is easier to learn than absolute memory selection when the address space is supplied by the environment. It reports faster threshold attainment for relative movement, while leaving learned value-indexed geometry untested.

3 A highly crude fixed-Φ proof of principle

  • The experiment supplies a hand-designed address map from memory-creation spatiotemporal state rather than learning Φ from value relations.It assumes physical and temporal proximity are useful indexing features.Source: §3 · p. 4
  • The task asks an agent to remember a cue while traversing a corridor, then select the corresponding object at a T-junction.A Transformer-XL policy is pretrained with context length 64, then fine-tuned on a randomized task with context length 5.Source: §3 · p. 5
  • The comparison tests relative head movement against Gaussian and categorical heads that select absolute memory positions.Each retrieval-head condition uses six seeds; the reduced-context no-head baseline uses three.Source: §3 · p. 5
  • 0.59M steps to the 0.95 threshold for relative movement, versus 0.94M for Gaussian and more than 2.0M for categorical absolute addressing.The comparison is on MiniGrid-Memory-S17Random; relative movement also produces trajectories close to the full-context reference with five context tokens.Source: Table 1 · p. 5, §3 · p. 5
  • A reduced-context baseline can encode cues through cue-dependent waiting or repeated corridor turns, so the experiment supports learning and behavioral efficiency, not a capability gap.The reported result is limited to incremental relative addressing in a single-retrieval task; it does not test the learned indexing objective, interaction-driven geometry, or low-drift hypothesis.Source: §3 · p. 5

4 Limitations

The proof-of-principle experiment leaves the framework’s central learned-geometry and repeated-retrieval claims untested. The section also identifies unresolved questions about representing context-dependent retrieval sets and training the proposed geometry.

4 Limitations

  • The experiment uses environment-provided geometry rather than learning Φ from the proposed objective, leaving the learned value-index experimentally unvalidated.It also retrieves only one piece of information, so it does not test sequential access as memories become useful.Source: §4 · p. 5
  • It remains unexplored how well fixed-dimensional geometry can represent the retrieval sets favored by Lλ(Φ).The open question includes whether positions can co-locate memories mainly in contexts where they are complementary, despite redundancy in other contexts.Source: §4 · pp. 5–6
  • The paper leaves open whether a single spatial arrangement can accommodate memory pairs that are redundant in some contexts but complementary in others.The section poses this as a question about whether their shared ε-ball membership can occur primarily in complementary contexts.Source: §4 · p. 6
  • The retrieval rule is left modular rather than canonical, and aligning head movement with construction of Φ under one principle requires a method modification.Source: §4 · p. 6
  • The paper gives the loss but does not clearly detail how to train Φ or describe the training procedure.Studying how sij varies with Φ may yield additional constraints that decompose the loss into easier-to-train terms.Source: §4 · p. 6

5 Conclusion

The paper frames long-context retrieval as persistent traversal through an address space shaped by retrieval utility, with local access balancing valuable combinations against unnecessary co-retrieval. Its MiniGrid experiment tests traversal only under fixed geometry; learning the value-indexed space and testing repeated low-drift retrieval remain next steps.

5 Conclusion

  • The proposed approach treats long-context retrieval as persistent traversal through an address space learned from retrieval utility.Source: §5 · p. 6
  • The objective makes valuable memory combinations locally accessible while charging for unnecessary co-retrieval, distinguishing complementarity from redundancy.It also shows that exact substitutes share the same interaction profile with other memory coalitions.Source: §5 · p. 6
  • The MiniGrid experiment tests only the traversal mechanism under fixed geometry; future empirical tests would learn the value-indexed space and evaluate repeated low-drift retrieval.Source: §5 · p. 6

A Open access to data and code

The authors provide experimental code and a reproduction script, with modest stated hardware and runtime requirements.

A Open access to data and code

B Declaration of LLM Usage

The authors report using general-purpose LLMs to assist with code development, theoretical formulation, and writing, while retaining responsibility for the paper’s claims.

B Declaration of LLM Usage

Related papers

Loading 2610.40545v1…