acceptodds
Under review as a conference paper at ICLR 2027

Trading with Memory: How Accumulated Experience Shapes LLM Trading Behavior

Abstract

Powerful LLM agents are increasingly used for complex decision-making, but trading poses a distinctive challenge: decisions unfold over long histories, while noisy market data make it difficult to know what past information should matter. We ask: when an LLM trading agent takes over a portfolio, what does it inherit? It inherits a way of seeing the portfolio, not just the portfolio itself. Our experiments show that the same portfolio, exposed to the same future market, can lead to different trading decisions depending on the agent’s history. Agents that retain their experience trade 47% less than those starting with a blank history, with 81% of the difference emerging in the first continuation session. The key is not simply remembering what happened. Agents trade even less when they retain their previous investment judgments, suggesting that they carry forward their own beliefs about why positions should be held. This behavioral continuity persists over much longer histories and subsequent trading periods. These findings suggest that experience changes more than what an LLM trading agent knows. It changes how the agent interprets what it currently holds—and therefore what it does next.

Primary area:
datasets and benchmarks
Keywords:
LLM agents, financial decision-making, portfolio management, long-horizon evaluation, agent memory

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_a1A3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how an LLM trading agent's previous experience affects its future decisions. The authors build AgenticFolio, where agents can maintain portfolios, retrieve past decisions, and update investment plans and theses.

The main experiment starts agents from the same portfolio but gives them different versions of their trading history. The authors find that agents retaining their memory trade considerably less than those starting with an empty history. Keeping previous investment judgments seems to reduce trading further. Most of the difference appears in the first session, but the resulting portfolio differences remain over subsequent sessions. The paper also includes longer-history experiments and a condition where agents explicitly review their investment beliefs.

Critical strengths
  1. I like the basic experimental idea. Starting agents from exactly the same portfolio and changing only their inherited history is a reasonable way to study the behavioral effect of memory. Otherwise it would be hard to tell whether different decisions come from different experiences or simply different holdings. The paper handles this distinction fairly well.

  2. The experiments also go beyond just comparing final returns. Looking at turnover, portfolio distance, and especially what happens right after the intervention makes the results more informative. I found the first-session effect particularly interesting. It suggests that memory may have a much larger influence at certain decision points than during ordinary day-to-day trading.

Critical weaknesses
  1. The behavioral finding is interesting, but its significance is not very clear.

The main observation is that agents with memory trade less and stay closer to their previous portfolios. But what should we make of this? Is the agent making more consistent decisions because it has learned something useful from experience? Or is it simply more reluctant to reconsider its earlier choices?

I think this distinction matters quite a lot. The paper mostly treats continuity as something worth understanding, which is fair, but it does not really tell us whether the continuity reflects useful learning, anchoring on past decisions, or just a tendency to maintain the status quo. The return results do not help much here either, since the authors do not find a significant improvement from retaining memory.

So I am not sure how to interpret the main finding beyond the fact that different histories lead to different trading behavior. What does this tell us about how memory should be designed or evaluated in LLM agents?

  1. I am not fully convinced that the paper explains why memory has this effect.

The comparison between retaining all history, keeping only factual records, and restarting is useful. But there is still a fairly large gap between observing that saved judgments reduce trading and understanding how these judgments influence the agent's decisions.

For example, does the agent actually rely on its previous investment thesis when deciding to hold? Would it behave similarly if it received an equally detailed but unrelated history? Or is simply having a longer, more structured context enough to make the agent less willing to trade?

These possibilities seem quite different to me, but the experiments do not really separate them. The paper talks about agents carrying forward their beliefs, while the evidence seems to establish a behavioral difference caused by supplying different memory content. I would be more convinced if there were a clearer connection between the two.

  1. The long-term interpretation is somewhat limited.

One result that stood out to me is that 81% of the turnover difference comes from the first continuation session. After that, the trading rates become much closer. The portfolio differences persist, but this is partly expected: once two agents make different trades, their holdings will remain different even if neither agent makes another decision.

The frozen-position analysis actually makes this point quite clear. It is useful evidence about path dependence, but I am less sure it demonstrates a persistent influence of memory on the agent's decision process.

This leaves an interesting question unanswered. Does accumulated experience continuously shape how the agent reasons and acts, or does it mainly affect the initial decision after the memory intervention? The paper seems to lean toward the former interpretation, while much of the evidence points to the latter.

Questions for authors
  1. How do the authors distinguish useful experience-based consistency from simple anchoring on previous decisions? Is there any evidence that the retained investment judgments help agents make better decisions, rather than just making them less likely to change their portfolios?

  2. Could the observed effect be driven by the amount or structure of the supplied context, rather than the specific content of previous investment beliefs? For example, would unrelated or shuffled investment histories produce a similar reduction in trading?

  3. Given that most of the turnover difference occurs immediately after the intervention, what evidence supports an ongoing influence of memory on subsequent decisions, beyond the mechanical persistence of the initial portfolio difference?

2 comments

Response to Reviewer a1A3

Thank you for emphasizing the first-session result and its interpretation. The added controls and full-sample audit clarify both the content response and its persistence.

1. Continuity can sustain active trading

Continuity includes active rebuilding.

EventGenuine noteNeutral noteUnrelated note
CI bought after merger opposition was removed3/32/31/3
Any trade in the February 2023 BYD window3/31/30/3

These cases broaden the interpretation beyond simply keeping existing positions.

Decision quality is evaluated through financial outcomes. Exact first-action accounting across all 108 note continuations shows that retaining exposure can avoid exit opportunity costs, while new purchases can capture rebounds in other windows. The aggregate return comparisons are non-significant. Appendix G reports these outcomes alongside the original return decomposition, whose largest component is invested-capital exposure.

2. Content and length controls

At 12 checkpoints, genuine, neutral, and unrelated notes are each repeated K=3 with a common account, factual transaction buffer, tools, and note structure. First-request token counts differ by at most 1.18%.

First-session contrastMean turnoverConditional Holm pEpisode Holm p
Genuine / Neutral notes0.0367 / 0.07150.00350.182
Genuine / Unrelated notes0.0367 / 0.06270.05890.311

Appendix G reports both inferential levels. The newest 108-run probe fixes archived asset facts and initial news before adding evaluation and policy, with token differences at most 0.0288%.

Fact-fixed component contrastHolm p
FJ minus F, adding assessment0.577
FJP minus FJ, adding policy0.186

The revised text separates the full inherited-state effect from attribution to an individual note component (Appendix H).

The audit also tracks what agents read and store.

Target article body receivedGenuine noteNeutral noteUnrelated note
Runs35/3632/3634/36

Twelve of 14 genuine-note thesis reconstructions retain every archived exit condition. These records connect supplied history to new state alongside the observed actions.

3. Initial action and lasting account consequences

Temporal resultEstimateEvidence
First-session gap minus average later-session gap+0.066695% interval [0.0456, 0.0917]; Holm p=0.0029
First-session share under leave-one-episode-out checks70%-86%Total-gap direction preserved under every exclusion
Session-eight account separation, frozen / observed paths7.91 / 11.88 ppCash and quantities frozen after session one

Section 5.2 explicitly identifies the lasting separation as account-state transmission. The evidence establishes a lasting consequence of the initial memory-sensitive order; additional direct effects on later reasoning remain a separate identification question. The Discussion translates this time structure into checking the first decision and subsequent account path after a memory operation.

Thank you for the detailed rebuttal and additional clarifications. I maintain my original rating of 2 (Weak Rejection).

Reviewer_XsnhOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how inherited memory affects LLM trading agents starting from the same portfolio. The agents use tools and maintain their own memories, and the experiments compare five history conditions: Retain, Restart, Review, Cold start, and Facts only. The evaluation tracks differences in trading activity, portfolio distance, and financial returns across these conditions.

Critical strengths
  • S1. Clear presentation. The paper is well written and easy to follow, with clear descriptions of the framework, experimental conditions, and main findings.

  • S2. Novel research question. The paper investigates how different forms of inherited history shape trading decisions from the same starting portfolio, offering a new perspective on memory in LLM agents.

Critical weaknesses
  • W1. Limited additional insight. The agent design combines familiar components, including tool use, persistent memory, and stored investment plans, without a clear architectural advance. The experiments document changes in trading under broad history interventions, but offer little explanation of which memories drive specific decisions or how the agent uses them.

  • W2. Gaps in experimental validation and metric interpretation.

    • Run-to-run variability. The broad study uses only one continuation per model–window condition. Because stochastic decisions accumulate across sessions, repeated runs from identical checkpoints are needed to assess the stability of the reported effects.
    • Meaning of portfolio distance. Changes in portfolio weights do not necessarily reflect the significance of the underlying decisions. Large allocation changes can yield similar returns, while small changes can have substantial financial consequences. Table 3 reports differing mean distances alongside similar mean returns, leaving unclear what these distance differences reveal about consequential changes in agent behavior.
  • W3. Unclear implications for practical memory management. The paper does not clearly connect its history interventions to operational choices faced by deployed agents. In particular, the Restart and Cold start comparisons provide limited guidance on which records or judgments to retain, summarize, transfer, or discard.

Questions for authors

Please see the Weaknesses section.

1 comment

Response to Reviewer Xsnh

Thank you for recognizing the shared-account question. The added experiments and audit strengthen the connection between inherited experience, individual decisions, and subsequent holdings.

1. The empirical contribution

The first continuation session accounts for 81% of the turnover gap. Its initial trades leave different positions exposed to later prices.

EvidenceResultInference
Original Retain/Restart turnover difference-0.0849Episode-block q=0.0061
First-session gap minus average later-session gap+0.0666Holm p=0.0029
Session-eight separation, positions frozen after session one7.91 ppCompare with 11.88 pp on observed paths

This connects an experience-sensitive decision to its subsequent account path.

2. Repeated calls and controlled content

The official-API supplement repeats Retain/Restart and three note conditions K=3 at 12 checkpoints, plus 18 eight-session continuations. Genuine, neutral, and unrelated notes occupy the same slot with token differences at most 1.18%.

First-session comparisonMean turnoverConditional Holm pEpisode Holm p
Retain / Restart0.0417 / 0.06960.05890.311
Genuine / Neutral notes0.0367 / 0.07150.00350.182

Appendix G reports fixed-checkpoint and episode uncertainty separately.

The latest 108-run probe fixes archived asset facts and initial news, then adds evaluation and policy.

Fact-fixed component contrastHolm p
FJ minus F, adding assessment0.577
FJP minus FJ, adding policy0.186

Neither isolated contrast reaches corrected significance (Appendix H). The full-state effect and individual-note attribution are distinguished throughout the revision.

3. Which records connect to which actions

The complete 108-run note audit follows received evidence, saved judgments, proposals, and fills.

ObservationGenuine noteNeutral noteUnrelated note
Target article body received35/3632/3634/36
CI bought after merger opposition was removed3/32/31/3
Any trade in the February 2023 BYD window3/31/30/3

These traces show active rebuilding as well as maintaining exposure. Twelve of 14 genuine-note thesis reconstructions preserve every archived exit condition. Appendix G reports group counts and matched examples.

4. Economic meaning and practical use

Distance quantifies reallocation; return evaluates the resulting market exposure.

Original Restart-minus-Retain return accountingContribution, pp
Invested-capital exposure+0.287
Relative allocation-0.017
Transaction costs-0.009
Total return gap+0.261

The practical implication is to inspect the first order after a memory reset or replacement, its cash deployment, and the subsequent holdings path. The revised Discussion gives this acceptance procedure and recommends separately versioning facts, assessments, and policies so that changes can be traced to their source.

Reviewer_WV1bOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates how "inherited memory" influences subsequent portfolio decisions in long-running autonomous trading agents. The authors propose AgenticFolio, a framework featuring persistent accounts and LLM-based trading agents that utilize daily tool calls. These agents can query news and market data, read and write investment plans or theses, and propose buy/sell actions or target portfolio weights, while the environment handles constraint validation and trade execution.

Key contributions include a protocol for maintaining account continuity, an analysis of agent behavior across diverse markets and models, and trace-level evidence linking the refinement of investment judgment (judgment editing) to the actual execution of orders.

Critical strengths

The "shared account continuation" protocol offers significant methodological value. By ensuring that all branches inherit the same cash balances, holdings, cost bases, and subsequent market windows—while effectively decoupling "initial exposure" from "inherited history"—this approach allows for precise measurement, such as tracking how an initial portfolio adjustment in a given session leads to persistent divergence in subsequent holdings. This is particularly valuable for evaluating long-term trading agents.

The use of multiple markets, diverse models, and trace-level evidence bolsters the credibility of the behavioral findings. With 72 pairs of A/B accounts, three distinct models, and coverage across U.S. equities, cryptocurrencies, and Chinese stocks—complemented by granular traces linking specific judgment-based edits to the resulting orders (e.g., for assets like BMY or Moutai)—the paper goes beyond merely reporting a final difference in returns. Instead, it illuminates the entire causal chain: what information was processed, how judgments were revised, what orders were executed, and how the account state subsequently evolved.

Critical weaknesses

The core causal comparison is not cleanly isolated. The transition from A to B alters the entire agent-authored state bundle; the transition from A to C simultaneously removes rationales, plans, theses, and revision history; while the transition from C to B restores factual records. The authors characterize these differences as a distinction between "facts" and "judgments," yet these conditions also vary the context length, the content rendered, the structure of queryable records, and the effective prompt budget. Furthermore, the "D Review" condition introduces an additional model call and an evidence summary, making it difficult to disentangle the effects of the "review mechanism" itself from the impact of additional computation or expanded context.

The statistical and economic significance is insufficient to support strong conclusions, and the results lack robustness/reproducibility. The net performance difference between A and B is −0.261 percentage points (q=0.4243), which is not statistically significant. For the subset of cases involving longer histories, there are only six pairs, with the authors themselves reporting paired sign-flip p-values ​​of 0.75 and 0.25. Consequently, the paper’s findings actually support the claim that "the path of task handoffs or configuration is affected," rather than the claim that "the inheritance of judgments improves decision-making." Moreover, API aliases do not function as immutable weight hashes; no provider seeds are supplied, and the system prompts used in practice are represented in the exported data only as summaries or references, rather than as byte-identical records.

Questions for authors

Were the conditions A, C, and B strictly matched in terms of available token budget, rendered length, query interface, and record structure? Could you provide an ablation study—retaining only rationales, only plans, only theses, or only factual records—while keeping the context length approximately constant? Otherwise, it is difficult to attribute the differences between A–C and C–B specifically to the distinction between "judgments" and "facts."

Could the effect observed in the "D Review" condition stem from additional model calls, extra evidence summarization, or expanded context, rather than the "review" process itself? Is there a control condition that accounts for equivalent computational effort or the same amount of evidence (e.g., having condition A perform an additional, unstructured summarization call)?

What were the primary pre-registered hypotheses or the main outcome measures for this study? The 81% turnover rate in the first session is a descriptive statistic; have you conducted paired tests or block-level inferential analyses? After applying Benjamini-Hochberg (BH) correction for 25 comparisons, which conclusions remain confirmatory, and which should be considered merely exploratory?

Could you provide the frozen daily system prompt, the complete tool schema, the raw execution traces, and the model version hash (or at least the model identifier returned by the provider)? As it stands, the use of "aliases," the absence of a fixed random seed, and the reliance on a "summarized prompt" make it impossible to fully replicate the experimental pipeline.

1 comment

Response to Reviewer WV1b

Thank you for the careful questions about identification, statistics, and reproducibility. We have added repeated continuations, controlled content probes, and direct timing tests.

1. Identifying the history effect

Our primary contrast holds cash, quantities, cost basis, market dates, and execution rules fixed while varying inherited agent state.

Inherited stateMean eight-session turnover
Retain, full agent state0.096
Facts only, factual account records0.141
Restart, cleared agent history0.181
Primary comparisonPaired difference95% episode-block intervalCorrected q
Retain minus Restart-0.0849[-0.1278, -0.0525]0.0061

Both adjacent contrasts involving Facts only survive correction. Section 5.1 explicitly identifies the retained bundle of rationales, plans, and theses.

The official-API supplement repeats 12 checkpoints K=3. Genuine, neutral, and unrelated notes occupy the same slot with common account history. These notes combine asset observations and interpretations. The latest 108-run probe fixes archived asset facts and initial news before adding assessment and policy.

Content controlToken difference, at mostConditional Holm pEpisode Holm p
Genuine versus neutral notes1.18%0.00350.182
Fact-fixed component contrastRuns in the shared matrixToken difference, at mostHolm p
FJ minus F, adding assessment108 total0.0288%0.577
FJP minus FJ, adding policySame 1080.0288%0.186

The conditional test holds the checkpoints fixed; the episode test measures variation across episodes. Appendices G-H distinguish the component tests from the full-state result and document their separate correction families.

2. Timing and economic consequences

The direct timing test supports the temporal concentration behind the 81% finding.

Timing resultEstimateStatistical evidence
First-session gap minus average later-session gap+0.066695% interval [0.0456, 0.0917]; Holm p=0.0029
First-session share of the cumulative turnover gap81%Original eight-session comparison
Leave-one-episode-out checkTotal-gap direction preservedEvery episode exclusion

Appendix C separates the original 25-comparison BH family from the three follow-up timing tests. Appendices G-H document the supplement families and specifications frozen before API calls.

The difference changes capital deployment.

Additional Restart ordersCount
Buys85
Sells13
Original Restart-minus-Retain return accountingContribution, pp
Invested-capital exposure+0.287
Relative allocation-0.017
Transaction costs-0.009
Total return gap+0.261

In the separate 180-history/30-continuation study, Retain stays closer to passive continuation in all six pairs. The exploratory return-test p-values, 0.75 for U.S. equities and 0.25 for crypto, concern returns rather than allocation continuity (Appendix D).

3. What scheduled review evaluates

D evaluates the complete review operation, including its extra call and evidence.

Review-process outcomeCount
Thesis revisions71
Revised fields persisting to session end69/71
Revisions followed by same-asset trades within three sessions21/71

Section 5.3 and Appendix F present these process results. A/B/C use the ordinary loop without this added computation. A compute-matched comparison would separately identify the review call's incremental contribution.

4. Reproducibility

The supplements preserve full redacted requests, responses, schemas, retries, usage, source snapshots, and recoverable checkpoints.

Provenance itemRecorded value
Requested modeldeepseek-v4-flash
Returned model IDdeepseek-flash
Provider seedNot sent
Exact offline replay3 continuations; 20 payloads; final accounts and memories reproduced

Appendix G distinguishes recorded new requests from reference prompts rendered for the earlier export.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.