acceptodds
Under review as a conference paper at ICLR 2027

STRATA: Spatiotemporal Alignment of Agentic Memory for Long-Horizon Preference Tracking

Abstract

Large language model assistants increasingly maintain relationships with users across months or years, yet most memory systems still treat personalization as a retrieval problem. This framing fails when a user's preferences change over time, and the turns most semantically similar to a query are no longer the correct value. We argue long-term memory should instead be framed as a state inference problem by computing the value of an attribute at a given time from its past states. We present STRATA, a memory architecture organized into three complementary layers (Spatial, Temporal, and Spatiotemporal) that create a single traversable graph via casual and relational edges. Preferences are stored as evolutionary chains, mapping the trajectory of its value over time. At inference time, an LLM agent traverses the memory graph with tool calls both within and across layers to intelligently retrieve relevant context. On PersonaMem, STRATA attains a new SOTA of 61.6% accuracy compared to six baselines, and a second-best on LongMemEval of 86.4% overall, driven by a 6.8-point margin on multi-session recall questions. Tool call analysis reflects the agent exhibiting emergent routing based on dataset and question type. Controlled ablations show the memory representation is crucial for preference tracking questions, and that agentic, query-conditioned traversal at inference time - a property unique to STRATA - drives meaningful gains.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
long-term memory, LLM agents, personalization, preference tracking, state inference, agentic retrieval, knowledge graphs, temporal reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_s1FCOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work introduces STRATA, an agent-centric memory architecture tailored to tracking user preferences over extended horizons. The authors reframe long-term personalization from traditional information retrieval into a latent state inference problem. To operationalize this, they build a three-tiered memory graph consisting of Spatial, Temporal, and Spatiotemporal tiers, capturing cross-session preference shifts through evolution chains, cross-layer connectivity, and policy-driven graph exploration.

Critical strengths
  1. Recasting long-term personalization as dynamic state estimation rather than static passage lookup is a well-motivated and meaningful perspective, particularly for settings where user intents shift over time. The spatiotemporal graph design is conceptually consistent with this framing.

  2. The pipeline is thoroughly structured, covering raw memory extraction, consolidation, multi-layered graph construction, inter-layer linking, and search traversal. The accompanying ablation studies demonstrate a complete and well-executed engineering effort.

  3. The selected evaluation suites—namely the Track Evolution, Revisit Reasons, and Aligned Recommendation tasks in PersonaMem—serve as direct tests for temporal and spatiotemporal representations. Performance degradation in ablation experiments demonstrates the functional utility of individual modules on these specific benchmarks.

Critical weaknesses
  1. The framework largely reads as an assemblage of established building blocks. Modules such as entity alignment, temporal sequencing, memory graphs, evolution trajectories, LLM-based memory consolidation, and autonomous graph walking are already well-trodden in the literature. While the authors organize these into a three-layer taxonomy, they do not articulate what genuinely novel algorithmic mechanism arises from this synthesis. Specifically, the boundary between evolution chains / spatiotemporal tiers and existing temporal knowledge graphs or event-centric memory systems remains nebulous.

  2. The manuscript lacks an exhaustive taxonomy-level comparison against prior memory paradigms. Because the primary contribution relies on reorganizing existing components, establishing rigorous and substantive distinctions from prior work is critical. Without a comprehensive related-work discussion, validating the claimed novelty is challenging.

  3. Graph synthesis relies on numerous hand-crafted rules (e.g., embedding similarity thresholds, heuristic entity matching, causal event criteria, and co-move linkages). The paper provides no systematic sensitivity analysis on these parameters, leaving it ambiguous whether performance stems from the proposed spatiotemporal formalism or simply from heavy trial-and-error hyperparameter tuning.

Questions for authors

Could you explicitly define the fundamental algorithmic innovation of STRATA? Beyond simply listing the architectural blocks, what sets STRATA apart from existing methods in temporal memory, graph retrieval, and personalization? Could the proposed evolution chains and agentic traversals be replicated by straightforward combinations of existing off-the-shelf tools?

Overall, I believe the main concerns of this paper are insufficient methodological novelty and the lack of convincing experimental evidence that STRATA provides substantial advantages over existing agentic memory methods. Several core components of the proposed framework are closely related to existing work on memory graphs, temporal memory, agentic retrieval, and personalization, yet the paper does not sufficiently explain what new algorithmic insight arises from their combination. Moreover, the improvement on PersonaMem is relatively small, while STRATA incurs substantially higher inference costs, making it difficult to justify the additional complexity based on the current empirical results.

2 comments
AuthorsOpenReview ↗

Response to Reviewer s1FC

Thank you for the expert and candid review. We apologize that this response arrives so close to the deadline and leaves you little time to consider it.

W1 / Q: The algorithmic innovation of STRATA. STRATA's novelty comes from three specific design decisions which, to our knowledge, have not been combined or tested before:

(1) Memory is partitioned by temporal behaviour, not by level of abstraction. Prior graph memories either use a single node type (A-Mem's notes, HippoRAG's phrases) or stack layers of increasing abstraction over the same content (Zep's episode / entity / community subgraphs, Mnemis's base and hierarchical graphs). STRATA instead separates records by how they change over time: persistent entities and facts (Spatial), dated events (Temporal), and preferences with their own history (Spatiotemporal), linked across tiers by shared evidence. To our knowledge, this is the first memory graph organized along the temporal axis with spatial "encoding" as well (hence the spatial/temporal naming).

(2) Reasoning happens at read time, over typed structure. Instead of a fixed retrieve-and-rerank procedure, a lightweight agent anchors in each tier and chooses which typed node to open, receiving it with its one-hop neighbors in the other tiers. This routing is unique to STRATA and largely emergent. With minimal routing rules in the prompt, the same gpt-4o-mini agent issues 60% of node opens on PersonaMem to the preference tier, against 16% on LongMemEval, where half go to events. To our knowledge, STRATA is the first system in which an agent chooses to navigate across tiers of a single memory graph.

(3) Append-only state with provenance. Preference (and fact) values are appended to chains instead of being overwritten, in contrast with overwrite-style memories (Mem0, LangMem, A-Mem's note rewriting). No other memory system uses temporally aligned chains as the main unit.

Q: Could this be replicated by combining off-the-shelf tools? We tested this via a no-graph control. The natural off-the-shelf assembly is a ReAct agent with dense retrieval over raw turns. We ran it with the same model, prompt, dense seed and per-item tool budget as STRATA:

BenchmarkSTRATA (3 runs avg)Same agent, no graphGain
LongMemEval-S (n=500)75.669.6+6.0
PersonaMem-32K (n=589)62.661.3+1.3

We are not aware of a prior memory-system paper that reports a budget-matched no-graph agent control; it isolates the architecture from the agent. The gain also concentrates where temporal organization should matter. On temporal-reasoning questions (n=133) STRATA scores 75.4 against 64.7 without the graph, and removing the event tier alone drops it to 61.7. On preference questions (n=30), removing the preference tier or the cross-layer edges drops accuracy from 67.8 to 46.7, and removing evolution chains or causal edges lowers it to 60.0. No ablation condition beats the full system on either benchmark.

W2: Taxonomy-level comparison. We added a table to Related Work comparing STRATA with fourteen other systems (including Zep, A-Mem, Mem0, MemGPT and Mnemis) on six axes, including how a changed value is handled, where the LLM reasons, and how memory is read. It makes the boundary with temporal knowledge graphs explicit: Zep attaches validity intervals to relation edges inside an abstraction hierarchy and reads with search plus reranking, while STRATA makes temporal behavior the organizing principle of the graph itself and reads with an agent.

W3: Sensitivity to hand-set rules. We added a table of hand-set values with their measured sensitivity: (a) causal-link threshold: accuracy on LongMemEval-S is 75.6 / 74.0 / 75.2 at cosine = 0.15 / 0.30 / 0.50, so the result is not a product of tuning it; (b) neighbours per relation (4) and the agent call limit (30) are non-binding; (c) cross-layer edges on vs off: 75.6 vs 71.8 on LongMemEval-S; (d) RAG baseline depth: flat from k=4 to k=32 (57.4–59.1), so it was not disadvantaged. The seed size and preference-merge threshold are not swept; we list this as a limitation.

Empirical advantage and cost. We re-ran every experiment with a gpt-4o-mini backbone, the same as every baseline, under a fixed inference budget (3,100-token answer input), on the full benchmarks (three runs). STRATA scores 62.59 on PersonaMem-32K (A-Mem 60.44, RAG 58.40) and 75.6 on LongMemEval-S, against 60.2 for the best system in All-Mem's table (arXiv 2603.19595), which uses the same backbone and embedder. STRATA now costs 0.027 per item on PersonaMem (0.023 one-time construction, 0.005 at query time), about 4 times A-Mem instead of 46x with Opus. We attribute the larger margin on LongMemEval to per-item context length; PersonaMem-32K items are one conversation of at most 32k tokens, while LongMemEval-S items span about 48 sessions and 104k tokens.

Thank you again for the thoughtful review. We appreciate your time and commitment.

Thank you for the response and additional experiments. I appreciate the matched-backbone evaluations, same-agent no-graph control, expanded discussion of prior work, and reduced computational cost. These additions address some of my original concerns and strengthen the evidence that STRATA’s structured memory pipeline is useful, particularly for temporal reasoning.

However, my central concerns remain. The response clarifies the architectural choices, but I remain unconvinced that it establishes a sufficiently distinct methodological contribution. The additional controls do not fully isolate the benefit of the proposed three-tier organization, and the improvements on the main preference-tracking benchmark remain modest relative to the added complexity and cost. The rebuttal therefore does not sufficiently resolve my concerns about novelty and empirical justification. I maintain my rating.

Reviewer_2FgWOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes STRATA, a long-term memory architecture that models evolving user preferences through a three-layer graph of entities, facts, events, and preferences. An LLM agent dynamically traverses temporal and relational links to reconstruct relevant user states at inference time. STRATA achieves 61.63% on PersonaMem and 86.4% on LongMemEval, with ablations examining the contributions of memory layers and agentic retrieval.

Critical strengths

S1. The paper addresses an important limitation of retrieval-based memory systems by explicitly modeling how user preferences evolve over time. The improvements on preference-related reasoning tasks support the usefulness of this representation.

S2. The ablations examine both memory structure and retrieval strategy. In particular, the performance degradation when replacing agentic traversal with a fixed retrieval policy provides evidence that adaptive memory navigation is beneficial.

Critical weaknesses

W1. Unclear gains under controlled comparisons. The paper claims a common GPT-4o-mini backbone, but STRATA's best result uses Claude Opus 4.6 as the retrieval orchestrator. With GPT-4o-mini as the orchestrator, STRATA achieves 58.23%, below A-Mem's 60.44%. Moreover, STRATA improves over A-Mem by only 1.19 points while incurring approximately 46× higher inference cost. These results make it difficult to establish a convincing performance advantage attributable to the proposed architecture.

W2. Limited differentiation from prior work. Zep already models temporally evolving facts using knowledge graphs, A-MEM supports evolving interconnected memories, and Mnemis combines semantic retrieval with structured graph traversal. While STRATA integrates explicit preference evolution chains with agentic traversal, its distinct contribution beyond these approaches is not sufficiently isolated. In particular, the paper lacks a controlled comparison demonstrating the added value of its temporal and causal structure over existing graph-based memory representations.

Questions for authors

Q1. Can the authors provide comparisons with A-Mem and Mnemis under matched orchestrators and inference budgets? How much of STRATA's improvement remains under these controls?

Q2. What is the specific advantage of STRATA's preference evolution chains and causal edges over existing temporal graph representations such as Zep? Can their individual contributions be isolated experimentally?

Q3. Why does STRATA underperform Full Context and RAG on Latest Preference questions (58.82% vs. 76.47% and 70.59%), despite its central motivation of inferring users' current preference states?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 2FgW

Thank you for the thoughtful review and the precise questions. We apologize that this response arrives so close to the deadline and leaves you little time to consider it.

W1: Matched Orchestrator. You are right that the submission's best result used a stronger orchestrator (Opus) than the baselines (gpt-4o-mini). First, after minimal prompt changes, we re-ran STRATA with gpt-4o-mini as the true model backbone (extractor, orchestrator and answer model), enforcing the same across every baseline along with a fixed inference budget (we capped the answer input at 3,100 tokens, which was the maximum input of all non-STRATA baselines which in this case was A-Mem). Concretely, each baseline gets a fixed: orchestrator, judge model, extraction model, inference budget, embedding model, and model temperature(s). On the full PersonaMem-32K (n=589), three runs give 62.48 / 62.31 / 62.99 (mean 62.59, sd 0.35).

SystemAcc.
STRATA62.59
A-Mem60.44
RAG58.40
LangMem / Mem0 / Full Context51.8 / 51.4 / 50.6

STRATA remains numerically highest under the new matched model.

W1: Cost. with the matched model STRATA costs 0.027 per item (0.023 one-time graph construction, 0.005 at query time) against A-Mem's 0.007, about 4 times the cost of A-Mem instead of the original 46x increase reported with the more-costly Opus orchestrator..

Q1: Mnemis. For LongMemEval-S, we have decided to drop the Mnemis numbers copied from Tang et al., as they did not use the same embedding model, so they are not fair comparators. Instead we cite All-Mem's table (https://arxiv.org/abs/2603.19595), which uses the same backbone (gpt-4o-mini) and embedder (all-MiniLM-L6-v2) as ours. STRATA maintains its position as the best system, scoring 75.6 on LongMemEval-S as the mean of 3 runs against All-Mem's highest of 60.2.

W2 / Q2. Isolating evolution chains and causal edges. We isolated the value of evolution chains and causal edges on the full benchmarks via ablation experiments reported below:

ConditionPersonaMem-32KLongMemEval-SLME temporal reasoning (n=133)LME preference (n=30)
Full system (mean of 3)62.5975.6075.467.8
− Evolution chains61.6374.6074.460.0
− Causal edges61.9773.6072.960.0
− Temporal (event) tier60.6169.20*61.7*63.3
− Spatial (fact) tier61.9771.40*72.950.0
− Preference tier61.2974.4072.946.7†
− Cross-layer edges60.7871.80*69.946.7†
Same agent, no graph61.2969.60*64.7*60.0

The full system remains the most accurate compared to all ablation experiments. Overall LongMemEval-S performance is not impacted significantly. However, if we focus on the category of questions that we expect to be most affected (e.g. Single-Session Preferences), we see that the removal of evolution chains and causal edges causes significant harm. The removal of the preference/spatiotemporal tier and cross-layer edges also cause the worst drops in performance.

Regarding PersonaMem, we attribute minimal changes in accuracy versus LongMemEval due to differences in context length: PersonaMem-32K items are one conversation of at most 32k tokens, while LongMemEval-S items span about 48 sessions and 104k tokens. Thus, PersonaMem questions are less likely to be affected by conversations with less numbers of edges/nodes. Regardless, the direction of all ablation results are still meaningful - no single ablation condition beats the full system in either benchmark.

On Zep differentiation: The previous ablation results support the explicit value of the temporal / causal features (e.g. preference evolution chains, or causal edges). Architecturally, Zep and Mnemis layer memory by level of abstraction (episodes, entities, communities), whereas STRATA separates records by temporal behaviour (entities and facts, dated events, preferences with history) and lets an agent choose which typed nodes to open. This is fundamentally different in both the memory organization layer, as well as the retrieval mechanism, where Zep relies on a fixed semantic retrieval + reranking approach with no LLM calls or agentic tools issued at query time. The temporal organization of STRATA (opposed to the semantic organization of Zep) is the fundamental separation.

Q3. Latest Preference. That category is the smallest by far, with only 17 questions. So, one question moves an accuracy reading by 5.9 points. We therefore attribute strong variation in this category's results to the benchmark's fault of poor category distribution. As a note, under the matched configuration STRATA scores 76.5 (three-run mean), equal to Full Context (76.5) and above RAG (70.6). In the revision, we temper our own conclusions drawn from this category given its small size.

Thank you again for the thoughtful review. We appreciate your time and commitment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to Reviewer jA4E

Thank you for the careful and detailed review. We apologize that this response arrives so close to the deadline and leaves you little time to consider it.

1. Model mismatch. You are correct, the claim of a common gpt-4o-mini backbone was not accurate given we used Opus as the orchestrator. We re-ran STRATA with gpt-4o-mini as the true model backbone (extractor, orchestrator and answer model), enforcing the same across every baseline along with a fixed inference budget (we capped the answer input at 3,100 tokens, which was derived from the maximum input of all non-STRATA baselines which in this case was A-Mem). Concretely, each baseline gets a fixed: orchestrator, judge model, extraction model, inference budget, embedding model, and model temperature(s). On the full PersonaMem-32K (n=589), three runs give 62.48 / 62.31 / 62.99 (mean 62.59, sd 0.35).

SystemAcc.
STRATA62.59
A-Mem60.44
RAG58.40
LangMem / Mem0 / Full Context51.8 / 51.4 / 50.6

STRATA remains numerically highest under the new matched model. Section 5.4 is rewritten around the new matched configuration, correcting any errors about wrong-direction sentences.

2. The no-graph agent control. We ran the following control: the same ReAct agent, model and prompt, the same dense seed, and iterative dense retrieval over raw turns as its only tool, with a per-item tool budget matched to STRATA's.

BenchmarkSTRATA (3 runs)ControlGain
LongMemEval-S (n=500)75.4 / 76.2 / 75.269.6+5.6 to +6.6
PersonaMem-32K (n=589)62.5 / 62.3 / 63.061.3+1.0 to +1.7

The graph adds a significant gain over agentic dense retrieval on both LongMemEval-S and PersonaMem-32K. We attribute the stronger gain on LongMemEval to context length: PersonaMem-32K items are one conversation of at most 32k tokens, while LongMemEval-S items span about 48 sessions and 104k tokens. Ablations show that the dense seed remains the largest component who's removal costs the most (removing it costs 6 points on PersonaMem and 15 on LongMemEval). We highlight this in the revision, but this alone does not discount the value of STRATA's graph. In particular, the RAG baseline is still inferior to both no-graph controls and full-benchmark baselines, showing that dense seed retrieval alone is not enough to close the gap, and that dense seed retrieval when combined with agentic traversal + a three-tiered graph offers the best value.

We also removed the fixed-traversal row, which cannot be compared fairly under a matched answer budget.

3. Scale and variance. All ablations are now on the full benchmarks, with the main configuration run three times. Run-to-run standard deviation is 0.35 (PersonaMem) and 0.53 (LongMemEval). The full system remains the most competitive, though we acknowledge the deltas in performance are not as pronounced (though this makes sense given the original numbers were skewed due to the small denominator in a 10% subset).

AblationPersonaMem-32KLongMemEval-S
Full system (mean of 3)62.5975.60
− Temporal tier60.6169.20
− Spatial tier61.9771.40
− Spatiotemporal tier61.2974.40
Temporal tier only60.6173.00
Spatial tier only60.1069.80
Spatiotemporal tier only61.6360.60
− Edges60.7871.80
− Evolution chains61.6374.60
− Causal edges61.9773.60
− Dense seed56.5460.40

4. LongMemEval protocol. We removed the baseline numbers copied from Tang et al., as they did not use the same embedding model, so they are not fair comparators. Instead we cite All-Mem's table (https://arxiv.org/abs/2603.19595), which uses the same backbone (gpt-4o-mini) and embedder (all-MiniLM-L6-v2) as ours. STRATA maintains its position as the best system, scoring 75.6 on LongMemEval-S against All-Mem's highest of 60.2.

5. Cost. With the matched model, STRATA costs 0.027 per item on PersonaMem (0.023 one-time graph construction, 0.005 at query time) against A-Mem's 0.007. This is actually about 4 times the cost, not 46x which was the original cost difference with Opus. The revised table reports construction, retrieval and answering separately for every system with updated values.

6. Causal threshold. You are right that 0.15 connects almost any pair: it links 95% of candidate preference–event pairs. Accuracy is insensitive to it (performance was 75.6 / 74.0 / 75.2 at cosine = 0.15 / 0.30 / 0.50 on LongMemEval-S), and removing causal edges entirely is not significant on either benchmark. We intended to use this edge as a nearest-prior-event link, not necessarily a semantic matching link - we refine our description of this edge to match, and no longer present it as causal evidence.

Thank you again for the thoughtful comments. The review strengthened the paper's claims substantially, and for the better.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.