acceptodds
Under review as a conference paper at ICLR 2027

When Does Typed Memory Help LLM Agents? A Subspace Perspective

Abstract

LLM agents increasingly rely on typed external memory, but it remains unclear when typing should outperform a strong flat retriever. We define a type as a subspace of a frozen representation space and typed memory as a complementary decomposition of stored content. At matched effective rank, typing provides no intrinsic capacity advantage over an unconstrained target-aware selector; its benefit must instead arise from inductive bias. When the target covariance has effective rank , blockwise spectral selection over label-free recovered subspaces incurs statistical dimension rather than , plus recovery and cross-block leakage costs. This gain belongs to the estimator rather than to physically separate stores: a flat retriever given the same blocks can obtain it without routing overhead, so store-specific benefits arise primarily from heterogeneous drift and per-type reliability. We instantiate these principles in , which learns write-time routing, complementary subspaces, per-type reliability and decay, and the effective number of types; provides the closed-form estimator covered by the analysis. Across seven text, image, and video benchmarks with a frozen encoder and matched rank , achieves versus for target-aware flat and for similarity-flat memory, with a benchmark-bootstrap improvement of over target-aware flat. Its gain is largest in the low-data regime: points at , at , and at . A label-free drift-heterogeneity index correlates with the cost of shared decay (Spearman , ). A risk-certificate heuristic, Type-If, also separates most negative controls from all seven benchmarks, although its decisions are sensitive to the convention constant. All empirical results use one backbone with a frozen encoder at matched effective rank.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, Agent memory, Typed memory, Subspace representations, Effective rank, Retrieval-augmented generation

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

The paper asks when splitting agent memory into types helps compared with a strong flat retriever. It treats types as subspaces of a frozen embedding space and argues that typing adds no capacity at a fixed rank budget. Any benefit instead has to come from easier estimation within recovered blocks or different retention policies for content that drifts at different rates. SOMA learns routing, subspaces, decay, and reliability, while a spectral variant is used for the analysis. The experiments report gains over target-aware flat memory, especially with limited task supervision, and use controls to separate blockwise estimation from the benefit of separate stores.

Critical strengths

I like the question and the effort to make the flat baseline strong. STRUCT-FLAT gets the recovered blocks too, so the comparison can separate the benefit of blockwise estimation from anything specific to separate stores. That's a useful distinction for someone deciding whether typed memory is worth adding to an agent.

The empirical work covers several plausible explanations for the gain: the full-objective K=1 control, alternative target constructions, the probe-budget sweep, and the drift intervention. I also took app. M into account. The ten-seed results give clearer evidence for the small store-specific increment, and there are task results on additional backbones. The aggregate task improvement has support within this protocol. What I struggle with is the claimed connection between the analysis and what the retrieval system actually does.

Critical weaknesses

My main concern is the link between store coverage and actual retrieval. How can the top-12 item span have only 0.03–0.08 slack relative to the full store in table 5, given the reported target effective rank? Under the common target covariance described in app. G, this seems inconsistent with the paper's own rank bound. If the calculation actually uses query-conditioned targets or pools retrievals across queries, the definition needs fixing. It would also leave open whether the subspace guarantees say much about the context returned to the agent.

I also don't see why target energy outside the recovered blocks should upper-bound the worst block-angle error in prop. 4(vi), p. 27. Capturing the combined span doesn't ensure that individual blocks were recovered. The claim that the measured recovery cost is conservative needs justification.

Questions for authors
  1. What exactly is measured in table 5? Please give one benchmark's store coverage, average top-12-span coverage, and target effective rank.

  2. What justifies the recovery-proxy upper bound in prop. 4(vi)? If it needs correction, do the TYPE-IF0 decisions change?

  3. How does prop. 5 handle near-ties between eigenvalues from different blocks at the pooled budget cutoff? Within-block gaps alone don't seem to prevent the allocation from changing.

  4. Does Stage A recover more than B basis directions? App. E specifies St(d, B), but algorithm 6 then selects B directions. As written, there appears to be nothing left to select.

Reviewer_a1uyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies when dividing LLM agent memory into different types is better than using a single memory store. The authors define each memory type as a subspace in the embedding space and argue that typed memory does not provide more storage capacity than a strong flat memory. Instead, its benefits come from learning with fewer labeled examples and handling different memory decay rates. Based on this idea, the authors propose SOMA, which learns how to organize memories into different subspaces, assign memories to them, and control their reliability and decay.

Critical strengths

The paper provides useful theoretical analysis and connects the benefits of typed memory to limited training data and different memory decay rates. The experiments cover several tasks and modalities, and the ablation studies help explain the contributions of different components. The finding that typed memory is more useful with fewer labeled examples is particularly interesting.

Critical weaknesses

The theoretical results do not directly explain the main SOMA model. The main theoretical guarantees apply to SOMA-2Sspec, which uses a fixed memory decomposition and a spectral estimator. However, the best experimental results come from joint SOMA, which learns its parameters together. The paper acknowledges this difference, but it is still unclear how much of the improvement in joint SOMA can be explained by the proposed theory.

The benefit of typed memory itself is relatively small. SOMA achieves 55.4%, compared with 54.2% for STRUCT-FLAT, which already uses the learned subspaces without separate memory stores. More importantly, the improvement from SOMA-2Sspec over STRUCT-FLAT is only 0.4 points, and its confidence interval includes zero. This makes it difficult to determine whether separate memory types provide a significant benefit beyond structured retrieval.

The proposed TYPE-IF0 rule has limited evidence for its ability to predict when typing helps. The rule recommends typed memory on all seven main benchmarks, but the paper reports that these benchmarks have very similar statistical conditions. Its decisions are also sensitive to the choice of a constant. It is unclear whether the rule can reliably distinguish between cases where typed memory helps and cases where it does not in more diverse settings.

Questions for authors

Can the authors provide more evidence that the theoretical findings also explain the improvements of joint SOMA, rather than only SOMA-2Sspec?

Since STRUCT-FLAT already achieves similar performance, what is the main advantage of using separate memory stores? Can the authors further isolate the effects of per-type decay and reliability?

How well does TYPE-IF0 perform on new tasks with different data distributions or memory drift patterns? How sensitive are its predictions to the constant used in the rule?

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

This paper studies when typed memory helps LLM agents compared with a flat retriever under the same rank budget. It defines memory types as subspaces and argues that their benefits come from easier estimation and different decay rates, rather than greater capacity. The proposed system, SOMA, improves performance across several benchmarks.

Critical strengths
  1. The paper separates several possible sources of improvement.

  2. SOMA improves over target-aware flat retrieval across the tested benchmarks. Appendix M adds ten-seed experiments and results on three backbones. These results support a small but consistent benefit under the tested setting.

Critical weaknesses
  1. The reported retrieval coverage seems inconsistent with its definition.

  2. Appendix G compares the coverage of the retrieved items with that of the full store and reports only 0.03 to 0.08 loss. However, the system retrieves 12 items, whose vectors span at most 12 dimensions. With the reported effective rank near 148, Equation (1) gives a coverage upper bound of (12/148). This is far below the reported store coverage of about 0.78 and cannot explain a loss as small as 0.04.

  3. Algorithm 6 allocates memory budget by comparing eigenvalues across blocks. However, the proof of Proposition 5 uses gaps within each block to argue that this allocation is stable. Large within-block gaps do not prevent eigenvalues from different blocks from changing order.

Questions for authors
  1. How exactly is the retrieval loss in Table 5 computed?

  2. Does Proposition 5 require an eigenvalue gap across blocks?

  3. How is target-aware flat trained and evaluated?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.