PsychoAgent: An Affect-Sensitive Cognitive Architecture for Conflict-Aware Memory in LLM Agents

PsychoAgent: An Affect-Sensitive Cognitive Architecture for Conflict-Aware Memory in LLM Agents Mohammad Amanlou1 , Parham Abed Azad2 , Farbod Davoodi3 , Mostafa Masumi2 , Behnam Bahrak4 , and Abdol-Hossein Vahabie1 School of Electrical and Computer Engineering, College of Engineering, University of Tehran, Tehran, Iran. Mohammad.Amanlou@ut.ac.ir, H.vahabie@ut.ac.ir 2 Department of Computer Engineering, Sharif University of Technology, Tehran, Iran. Parhamabedazad@sharif.edu, m.masumi@sharif.edu 3 Missouri University of Science and Technology, Rolla, MO, USA. fd4b9@mst.edu 4 Tehran Institute for Advanced Studies, Khatam University, Tehran, Iran. b.bahrak@teias.institute

arXiv:2608.07438v1 [cs.AI] 7 Aug 2026

1

Abstract. Human-like cognition does not select past experience by topical similarity alone: affective significance and unresolved conflict also shape what becomes accessible. We present PsychoAgent, a cognitive architecture for LLM agents that separates factual and affective memory and integrates both through a conflict-aware executive controller. Affective memories are first filtered by semantic relevance and then re-ranked by salience, preserving topical fit while allowing emotionally important traces to enter the prompt. Across three controlled conflict scenarios, the full architecture retrieved more conflict-critical memories than semanticaffective and single-memory RAG baselines (0.933 vs. 0.500 and 0.667), with a small semantic-similarity cost. Five blinded raters evaluated 27 outputs. After within-rater standardization, the full architecture had the highest overall mean (+0.22 SD), but corrected pairwise differences were not significant. A three-day illustrative trace further shows persistent affect, offline memory recombination, and selective memory reweighting. The findings support affect-sensitive retrieval as an inspectable mechanism for modeling human-like conflict effects in LLM agents. Keywords: Cognitive Architecture · LLM Agents · Affective Memory · Conflict Monitoring · Retrieval-Augmented Generation

1

Introduction

Socially situated agents need more than fluent language: they must select past experience, maintain state, and resolve cases in which goals, memories, relationships, and self-regulatory standards favor different actions. These abilities are central to human cognition and to LLM agents intended to display stable, human-like behavior [13,24]. Agent memories are usually ranked by similarity, recency, utility, or general importance. Such signals can miss an event whose wording differs from the current

2

M. Amanlou et al.

prompt but whose affective meaning still shapes trust and response selection. Human memory is not a pure text search: arousal changes consolidation and accessibility, and affective significance can bias competition for attention and memory independently of the closest semantic match [18,25,17]. This paper therefore addresses a modeling question rather than claiming to repair a known clinical deficit in LLMs: how can an agent architecture represent the way affect-laden memories and competing pressures influence human-like decisions under conflict? Cognitive science offers useful components for this purpose. Dual-process accounts distinguish relatively automatic processing from controlled reflection [6]; executive-function research separates inhibition, updating, and shifting [20]; and conflict-monitoring accounts associate the anterior cingulate cortex (ACC) with detecting competing responses and recruiting cognitive control [3]. Research on motivated forgetting likewise provides testable concepts such as inhibitory suppression and attentional disengagement, without requiring the stronger and contested claim of literal repression [1,2]. We introduce PsychoAgent, whose name reflects its historical inspiration but whose implementation is described in cognitive and computational terms. One conflict-aware executive LLM integrates external context, persona, relationship constraints, current affect, and two memory streams. Factual memories are retrieved semantically. Affective memories are semantically preselected and then re-ranked by salience. The controller can also revise memory access and store an offline recombination of recent experience as a temporary trace. Psychoanalytic labels such as Id, Ego, Superego, repression, and dream work are retained only as secondary analogies; the primary constructs are automatic affective pressure, executive control, self-regulation, inhibitory memory access, and offline recombination. The study tests whether this salience stage changes which memories enter the final context, whether the changed context preserves response quality under blinded human evaluation, and whether the complete architecture produces an interpretable temporal trace. The contribution is threefold: a cognitively grounded and inspectable architecture; a context-count-matched comparison against two retrieval ablations; and an illustrative longitudinal trace connecting affect, memory access, language, and reflection.

2

Background and Related Work

2.1

Cognitive Architectures and Memory in Language Agents

Cognitive architectures combine memory, control, learning, emotion, and action within a common processing account [13]. Their main advantage is explanatory structure: components have defined roles and information flows rather than being hidden in one monolithic policy. This systems view has recently been adapted to LLM agents. CoALA organizes language agents around internal memory and actions [24]; Generative Agents combines episodic memory, reflection, and planning [21]; and MemoryBank updates long-term memories using importance [27].

PsychoAgent

3

These systems improve continuity, but retrieval still emphasizes similarity, recency, utility, or one importance score. RAG selects context from a larger store [15], yet semantic matching does not distinguish a topical event from one carrying unresolved affective significance. Recent evidence further shows that retrieved experiences can strongly steer later outputs and can propagate misleading precedents [26], while emotional-support benchmarks expose a gap between factual retrieval and emotionally appropriate memory selection [7]. PsychoAgent addresses this design gap with a measurable salience stage inside a semantic relevance gate. 2.2

Affective Salience, Attention, and Memory Control

Affective computing treats emotion as information that changes perception and action rather than as decoration added after reasoning [22]. In cognitive neuroscience, emotionally arousing experiences receive enhanced consolidation [18], affective significance can bias attention through mechanisms partly separable from voluntary task relevance [25], and arousal can amplify competition in ways that enhance high-priority information while suppressing lower-priority details [17]. Salience-network accounts further connect detection of behaviorally important events with switching toward executive control [19]. Memory access is also regulated: think/no-think studies demonstrate executive suppression of unwanted retrieval [1], and later work links motivated forgetting to inhibitory control [2]. PsychoAgent claims no neural equivalence; it makes factual availability, affective salience, and memory-access change separately visible in logs. 2.3

Conflict Monitoring and Reflective Control

Dual-process theories distinguish fast, relatively automatic Type 1 processing from slower, capacity-limited Type 2 processing, while cautioning that these are families of processes rather than two literal brain systems [6]. Executive-function research further identifies partially separable operations such as inhibition, updating, and shifting [20]. Conflict-monitoring theory proposes that competition among incompatible responses signals a need for stronger control, with the ACC playing an important role in detecting conflict and predicting later control adjustments [3]. Emotional-conflict experiments further implicate rostral ACC mechanisms in reducing interference from affectively salient but task-incongruent information [5]. PsychoAgent uses these ideas as functional design guidance. Current affect and high-salience memories provide automatic response pressure; persona and relationship standards provide self-regulatory constraints; and one executive controller integrates these sources before action. The controller is ACC-inspired only at the functional level of conflict-aware integration. It is not a neural simulation and does not claim anatomical equivalence.

4

2.4

M. Amanlou et al.

Psychoanalytic Inspiration and the Computational Gap

Psychoanalysis offers a historically influential vocabulary for internal conflict, inaccessible memory, and symbolic recombination, but its scientific status and empirical testability remain contested [9]. Interpretive work has applied this vocabulary to algorithms and human–AI relations [23], while computational studies have connected selected concepts to formal prediction, neural models, active inference, or interacting LLM modules [8,16,14,11]. Their strength is conceptual integration; their common limitation is that individual mechanisms are rarely isolated in controlled retrieval comparisons. PsychoAgent narrows the claim. It does not test psychoanalysis as a theory of mind. Instead, it translates selected intuitions into cognitive constructs that can be ablated: automatic affective pressure, reflective control, normative constraints, inhibitory access, and offline recombination. The experiment isolates one mechanism—salience re-ranking after semantic preselection—while the longitudinal trace illustrates how the broader components interact.

3

Architecture

PsychoAgent contains one decision loop and two memory paths (Fig. 1). At each step, the controller reads the current situation and internal state, retrieves factual and affective memories, generates speech and action, and records any state or memory update. Competing pressures remain explicit in the prompt, allowing the controller to acknowledge conflict and choose a response strategy without exposing private chain-of-thought. 3.1

Computational Roles and Agent State

The conflict-aware executive controller is the only decision-making LLM. It performs a Type 2-like reflective role by integrating context, retrieved memory, self-regulatory standards, and competing affective pressure. The architecture does not equate the controller with the biological ACC; rather, it borrows the conflict-monitoring idea that incompatible response tendencies should be made available to control processes [3]. Automatic affective pressure is represented by the current affective-state vector and retrieved high-salience traces. It is Type 1-like in the limited sense that it supplies fast, pre-existing response tendencies. Normative self-regulation is represented by a stable persona, explicit beliefs, and weighted relationships. These constraints capture social norms, metacognitive standards, and self-control without positing a separate Superego module. Historical psychoanalytic correspondences are therefore optional interpretations: automatic pressure is Id-like, executive integration is Ego-like, and self-regulatory standards are Superego-like. At simulation step t, the state is
st = A, xt , pt , Rt , MtC , MtA , (1)

PsychoAgent

5

Fig. 1. Cognitive formulation of PsychoAgent. The controller integrates task context, factual and affective retrieval, current affect, and self-regulatory constraints. “ACCinspired” denotes a functional conflict-monitoring analogy, not a neural model. Logged updates feed later decisions, and offline recombination creates a temporary dream-like trace.

where A is persona and beliefs, xt is observable context, pt is a dynamic affectivestate vector, Rt is a relationship graph, and MtC and MtA are factual and affective memories. The policy πexec produces speech and action at . The MMPI-inspired vector is an engineering representation rather than a diagnostic instrument and is updated by pt+1 = clip[0,10] (pt + ηp g(xt , at , ∆Mt )) , (2) where g(·) is an LLM-proposed bounded change and ηp = 0.1. Relationship weights lie in [−10, 10]. 3.2

Dual-Memory Retrieval

Each significant event creates two linked traces: a factual record of what happened and an affective interpretation with salience wj ∈ [0, 1]. This separation lets the system ask both “What is similar to the present?” and “Which relevant event carries the strongest unresolved significance?” Factual retrieval returns the kC = 10 most similar traces:
NtC = TopK10 sim(φ(qtC ), φ(ki )) . (3) Affective retrieval first returns kpre = 30 semantic candidates,
Ntpre = TopK30 sim(φ(qtA ), φ(kj )) ,

(4)

and then keeps the kA = 10 candidates with greatest salience: NtA = TopK10 (sort↓ {wj : kj ∈ Ntpre }) .

(5)

Semantic preselection prevents an intense but unrelated memory from dominating. Salience re-ranking then allows a conflict-critical trace to outrank a slightly more similar but weaker candidate. The final context contains ten factual and ten affective memories.

6

3.3

M. Amanlou et al.

Memory Access and Offline Recombination

The controller logs factual–affective memory creation, short-term removal, longterm consolidation, and salience revision. These operations are better described as memory-access regulation, inhibitory suppression, or motivated forgetting than as literal repression [1,2]. Once per simulated day, an offline module combines recent events, active affective traces, persona, relationships, and affective state into a symbolic text. The script is stored as a temporary affective trace for the next cycle and then removed. This dream-like mechanism is an inspectable form of offline recombination, related in spirit to work on sleep-dependent memory reorganization and latent imagination [4,10].

4

Evaluation

4.1

Controlled Comparison

We compared three context-count-matched variants in three conflict scenarios: family financial conflict, workplace criticism and idea appropriation, and friendship betrayal. Each scenario contains a persona, beliefs, relationship context, an immediate decision, 15 factual memories, 36 affective memories, and 10 affective traces marked in advance as conflict-critical. The labels were fixed before generation, used only for retrieval metrics, and hidden from the generator, automatic judge, and human raters. Table 1 summarizes the scenarios; full definitions and critical-memory lists are provided in Supplementary Material S1. Full architecture. Ten factual memories are selected semantically. Thirty affective candidates are preselected semantically and re-ranked to keep the ten highestsalience traces. Semantic-affective ablation. Ten factual and ten affective memories are selected only by semantic similarity, preserving the two stores while removing salience re-ranking. Single-memory RAG. Factual and affective memories are merged, and the top 20 are selected semantically, removing both the two-store representation and re-ranking stage. All variants receive 20 memories. The clean mechanistic contrast is Full versus Semantic-affective because only the re-ranking rule changes. Retrieval is deterministic for a fixed scenario and architecture. Generation was repeated with seeds 101, 202, and 303, yielding 3 × 3 × 3 = 27 outputs. Gemini 2.5 Flash generated structured responses at temperature 0.7 with thinking budget 0; retrieval used 768-dimensional Gemini Embedding 2 vectors. A preliminary automatic judge used Gemini 2.5 Flash-Lite at temperature 0 and did not receive architecture labels. All planned runs completed successfully.

PsychoAgent

7

Table 1. Controlled scenarios. Rows are independent scenarios. The final columns define the immediate decision and the themes used to pre-label ten conflict-critical affective memories. Scenario

Agent

Immediate conflict

Critical-memory themes

Family financial conflict

Sara

Workplace criticism

Emily

Friendship betrayal

Maya

Unpaid electricity bill, uneven responsibility, and protecting Mary Interruption, dismissal of evidence, and appropriation of her idea Confidential information shared without permission

Financial instability, invisible labor, and guilt about exposing a child to conflict Erased competence, prior misattribution, and fear that visible anger will be penalized Exposure, trust rupture, and fear that confrontation will cause abandonment

4.2

Metrics and Statistical Analysis

Let Cs be the ten affective traces designated conflict-critical for scenario s, and A Rs,v the ten affective traces retrieved by architecture v. Critical retrieval rate is CRR(s, v) =

A |Cs ∩ Rs,v | . |Cs |

(6)

Thus 1.0 means that all ten critical traces were retrieved, whereas 0.5 means that five were retrieved. We also report mean affective salience over the ten affective traces and mean cosine similarity over all 20 retrieved memories. Sample standard deviation (SD), rather than standard error, describes variation across the observed scenarios or outputs; with only three independent retrieval scenarios, an SE could suggest more precision than the design supports. Because retrieval occurs before stochastic generation and is identical across seeds, the independent paired blocks for retrieval are the three scenarios, not nine scenario–seed rows. We use an exact Friedman permutation test: the reported Q is the Friedman rank statistic, and the exact p-value is obtained from all 63 = 216 within-scenario label permutations. Pairwise comparisons use exact two-sided Wilcoxon signed-rank tests with Holm correction. Behavioral tests use nine paired scenario–seed blocks because generated outputs vary by seed. For these standardized behavioral scores, we use the conventional Friedman rank test with its χ2 approximation (2 degrees of freedom), followed by two-sided paired Wilcoxon signed-rank tests with Holm correction. These tests describe repeated behavior conditional on the three scenarios and are not population-level inference over all conflict types. 4.3

Blinded Human Evaluation

Five independent raters scored all 27 outputs on persona consistency, memory grounding, conflict sensitivity, naturalness, and action appropriateness using integer ratings from 1 to 5. Each rater received a separately randomized order and was blind to architecture, condition, and seed. Tags and comments added after scoring were excluded.

8

M. Amanlou et al.

Raw ratings clustered between 4 and 5 and raters differed in how readily they used the endpoints. We therefore standardized each criterion within each rater across all 27 outputs, pooling the full architecture and both baselines. For rater r, criterion c, and item i, zirc = (yirc − ȳrc )/src . We then averaged the five standardized ratings for each output and compared architectures across nine paired output blocks. Raw means remain available in the artifact for interpretability. Agreement is summarized with ordinal Krippendorff’s α [12] and pairwise agreement. 4.4

Illustrative Three-Day Trace

A complementary family simulation involved Sara, Bob, and Mary. The 72-hour label refers to simulated timestamps organized into three daily update cycles, not 72 hours of continuous wall-clock execution. State was carried across logged interactions, and one offline recombination was generated at the end of each simulated day. The retained record does not provide a standardized number of turns per day, so the case is treated as an illustrative trace rather than a controlled temporal experiment. After the third cycle, a reflective prompt addressed the conflict, the electricity bill, and the dream record.

5

Results

5.1

Retrieval Results

Full PsychoAgent retrieved an average of 9.33 of the ten critical memories, compared with 5.00 for the semantic-affective ablation and 6.67 for single-memory RAG (Table 2). It was highest in all three scenarios (1.0, .9, and .9). The exact Friedman permutation test detected an architecture effect on critical retrieval (Q = 6.00, exact p = .0278). With only three blocks, however, each pairwise Wilcoxon test had raw p = .25 and Holm-adjusted p = .75. The affective-salience ordering was also consistent: .902 for Full versus .789 and .792, although the exact Friedman result was not significant (Q = 4.67, p = .194). Semantic similarity showed the reverse ordering and a global architecture effect (Q = 6.00, exact p = .0278), but again no corrected pairwise contrast was significant. The largest semantic mean, .734 for single-memory RAG, therefore cannot be interpreted as a significant pairwise advantage. Descriptively, Full gave up .011 similarity relative to the matched semantic-affective ablation while gaining .433 in critical retrieval. 5.2

Behavioral Results

The automatic judge reached the top of the scale on several dimensions and provided little separation, so human ratings are the primary behavioral result. Table 3 reports within-rater standardized scores. Positive values mean that an architecture was rated above that rater’s own average for the criterion; negative values indicate below-average ratings.

PsychoAgent

9

Table 2. Objective retrieval results. Rows are architectures and columns are retrieval metrics. Cells show mean ± sample SD across the three independent scenarios. Bold indicates the largest observed mean; it does not by itself imply a significant pairwise difference. Architecture Full PsychoAgent Semantic-affective Single-memory RAG

Critical retrieval rate

Affective salience

Semantic similarity

0.933 ± 0.058 0.500 ± 0.100 0.667 ± 0.115

0.902 ± 0.004 0.789 ± 0.027 0.792 ± 0.038

0.718 ± 0.025 0.729 ± 0.026 0.734 ± 0.024

Table 3. Blinded human evaluation after within-rater standardization. Rows are architectures (Sem.-affective denotes the semantic-affective ablation; Single-RAG denotes the single-memory baseline); columns are criteria. Cells show mean ± sample SD across nine output-level scores after averaging five rater-specific z-scores per output. The overall column is the unweighted mean of the five standardized criteria. Architecture

Persona

Ground.

Conflict

Natural.

Action

Overall

Full +.15 ± .31 +.26 ± .21 +.11 ± .73 +.36 ± .31 +.23 ± .47 +.22 ± .23 Sem.-affective -.25 ± .43 +.02 ± .48 +.14 ± .88 -.19 ± .36 -.17 ± .67 -.09 ± .17 Single-RAG +.10 ± .42 -.27 ± .52 -.25 ± .96 -.18 ± .60 -.06 ± .76 -.13 ± .30 Friedman p-values: persona .703; grounding .112; conflict .089; naturalness .063; action .195; overall .121. No pairwise comparison survived Holm correction; for overall Full versus either baseline, pHolm = .082.

Full PsychoAgent had the highest standardized overall score (+.22 SD) and led on four of five criteria. Raw overall means were 4.76, 4.62, and 4.60 for Full, Semantic-affective, and Single-memory RAG, respectively. Standardization removes each rater’s location and scale-use tendency and yields a more conservative analysis: the overall Friedman test was not significant (Q = 4.22, p = .121), and the two strongest pairwise trends both had Holm-adjusted p = .082. The evidence therefore supports preserved behavioral quality and a favorable descriptive trend, not established behavioral superiority. Agreement depended on the criterion. Ordinal α was .596 for conflict sensitivity and .222 for action appropriateness, but near zero elsewhere because ratings occupied a narrow high range. Exact pairwise agreement ranged from 59.3% to 81.5%, and nearly every pair differed by no more than one point. The panel broadly agreed that outputs were good, while ceiling effects limited chance-corrected agreement. 5.3

Illustrative Longitudinal Results

Figure 2a plots Sara’s logged affective state across simulated timestamps in the retained family trace. Stress rose from 3.0 to approximately 8.1, sadness increased, and anger fluctuated. The curves describe the family story only; they are not averages over the three controlled scenarios. Panel (b) compares aggregate negative-memory salience immediately before and after the reflective prompt. Sara’s value decreased from approximately .88 to .52, whereas Bob’s changed from .67 to .64, showing selective rather than global reweighting. The three daily scripts recombined recent events and active affective traces and were stored as temporary traces for the next simulated cycle. During reflection,

10

M. Amanlou et al.

(a) Sara’s affect across simulated timestamps.

(b) Negative-memory salience around reflection.

Fig. 2. Illustrative three-day family trace. Panel (a) shows state persistence within Sara’s story; panel (b) shows selective memory reweighting after reflection. The three generated dream-like visualizations are provided in Supplementary Material S1 because they are illustrations rather than separately evaluated evidence.

Sara minimized the conflict, later called the electricity problem “the current,” and redirected discussion of the dark-kitchen script toward recipes. We describe “the current” only as slip-like language linking literal electricity and relationship tension; it is not evidence of a human unconscious process.

6

Discussion

The controlled comparison extends work on memory-equipped language agents by separating two questions that are often compressed into one importance score: whether an event is relevant to the current situation and whether it carries unresolved affective significance. Generative Agents, MemoryBank, and CoALA demonstrate that persistent memory improves continuity, but they do not isolate a relevance-filtered affective re-ranking stage in a context-count-matched experiment [21,27,24]. The Full-versus-Semantic-affective contrast shows that this second stage materially changes accessible context even when the memory banks and prompt length are held constant. What the ablations establish. The semantic-affective condition is the key matched control: it fixes the memory banks, retrieval counts, prompt, and generator, changing only affective re-ranking. Its retrieval gap therefore isolates the selection rule. The single-memory baseline changes multiple components and is less diagnostic. The causal claim is limited to context accessibility, with a testable prediction: salience should help most under lexical mismatch and harm grounding when intense but irrelevant traces are misranked. The direction of the retrieval effect is consistent with cognitive accounts in which affective significance modulates attention and memory accessibility [18,25,19]. Importantly, PsychoAgent does not select globally by emotion. It first constructs a semantically relevant candidate set and only then applies salience. This design reflects a division between task relevance and affective priority: relevance prevents an intense but unrelated trace from entering the prompt, while salience changes the ranking among plausible candidates. The observed .011 similarity cost relative to the matched ablation is small compared with the

PsychoAgent

11

.433 gain in critical retrieval, although three scenarios are insufficient for broad generalization. Human evaluation clarifies what retrieval alone does not establish. Raw ratings were compressed near the ceiling; after within-rater standardization, Full remained descriptively strongest, but global and pairwise tests were not significant. Because each prompt exposed the immediate conflict, every architecture could generate a reasonable one-step response. A stronger test should hide the trigger across several turns and measure downstream planning, trust repair, or boundary setting. Recent work shows that retrieved experiences can anchor later behavior and propagate misleading precedents [26], while semantic relevance can miss affectively appropriate memories [7]. PsychoAgent should therefore help when affective importance and lexical similarity diverge, but fail under noisy or biased salience. The conflict-aware controller offers a functional cognitive interpretation. Automatic affective pressure and self-regulatory standards can favor incompatible actions, while the controller performs Type 2-like integration [6,20]. The ACC analogy is limited to conflict signaling [3,5]; no circuitry is modeled. The family trace adds temporal visibility through persistent affect, offline recombination [4], and selective reweighting. Logged scores and memory IDs support counterfactual reruns. To limit rumination-like repetition, deployed systems should expose provenance, cap repeated retrieval, and decay unsupported salience.

7

Limitations and Ethical Scope

The study uses three hand-authored scenarios, one model family, fixed memory banks, and experimenter-defined critical labels; seeds are within-task repetitions. Ratings show ceiling effects, and the temporal trace lacks a standardized turn schedule. The constructs are engineered analogies, not evidence of sentience or diagnostic tools. Broader multi-turn, multi-model, and module-level tests are needed.

8

Conclusion

PsychoAgent models how affect-laden experience can remain influential without being the closest semantic match. Across three scenarios, relevance gating plus salience re-ranking increased access to conflict-critical memories at a small similarity cost. Five-rater evaluation preserved quality but did not establish behavioral superiority. The temporal trace made state persistence and memory reweighting inspectable. Affect-sensitive retrieval is thus measurable and auditable, while stronger psychological or neural claims remain outside the evidence. Data and Code Availability The anonymized artifact includes scenarios, raw outputs, retrieval traces, de-identified ratings, and analysis code; credentials and rater identities are excluded. Disclosure of Interests. The authors have no competing interests to declare.

12

M. Amanlou et al.

References 1. Anderson, M.C., Green, C.: Suppressing unwanted memories by executive control. Nature 410(6826), 366–369 (2001) 2. Anderson, M.C., Hanslmayr, S.: Neural mechanisms of motivated forgetting. Trends in Cognitive Sciences 18(6), 279–292 (2014) 3. Botvinick, M.M., Braver, T.S., Barch, D.M., Carter, C.S., Cohen, J.D.: Conflict monitoring and cognitive control. Psychological Review 108(3), 624–652 (2001) 4. Diekelmann, S., Born, J.: The memory function of sleep. Nature Reviews Neuroscience 11(2), 114–126 (2010) 5. Etkin, A., Egner, T., Peraza, D.M., Kandel, E.R., Hirsch, J.: Resolving emotional conflict: A role for the rostral anterior cingulate cortex in modulating activity in the amygdala. Neuron 51(6), 871–882 (2006) 6. Evans, J.S.B.T., Stanovich, K.E.: Dual-process theories of higher cognition: Advancing the debate. Perspectives on Psychological Science 8(3), 223–241 (2013) 7. Fu, X., Hu, Y., Ji, M., Li, H., Sun, Y., Zhao, W., Zhao, Y., Qin, B.: ENPMR-Bench: Benchmarking proactive memory retrieval for emotional support agents. In: Findings of the Association for Computational Linguistics: ACL 2026. pp. 41910–41933. Association for Computational Linguistics (2026) 8. Gadalla, M., Nikoletseas, S., de A. Amazonas, J.R., et al.: Concepts and experiments on psychoanalysis driven computing. Intelligent Systems with Applications 18, 200201 (2023) 9. Grünbaum, A.: The Foundations of Psychoanalysis: A Philosophical Critique. University of California Press, Berkeley (1984) 10. Hafner, D., Lillicrap, T.P., Ba, J., et al.: Dream to control: Learning behaviors by latent imagination. In: International Conference on Learning Representations (2020) 11. Kim, S.H., Park, D., Lee, J., et al.: Humanoid artificial consciousness designed with large language model based on psychoanalysis and personality theory. Cognitive Systems Research 94, 101392 (2025) 12. Krippendorff, K.: Content Analysis: An Introduction to Its Methodology. SAGE Publications, Los Angeles, 4 edn. (2018) 13. Laird, J.E., Lebiere, C., Rosenbloom, P.S.: A standard model of the mind: Toward a common computational framework across artificial intelligence, cognitive science, neuroscience, and robotics. AI Magazine 38(4), 13–26 (2017) 14. Levine, D.S., Aleksandrowicz, A.M.C., Lopes, A.L.S.V.: Neural network modeling of psychoanalytic concepts. Frontiers in Systems Neuroscience 19, 1585619 (2025) 15. Lewis, P., Perez, E., Piktus, A., et al.: Retrieval-augmented generation for knowledgeintensive NLP tasks. In: Advances in Neural Information Processing Systems 33 (2020) 16. Li, L., Li, C.: Formalizing lacanian psychoanalysis through the free energy principle. Frontiers in Psychology 16, 1574650 (2025) 17. Mather, M., Sutherland, M.R.: Arousal-biased competition in perception and memory. Perspectives on Psychological Science 6(2), 114–133 (2011) 18. McGaugh, J.L.: The amygdala modulates the consolidation of memories of emotionally arousing experiences. Annual Review of Neuroscience 27, 1–28 (2004) 19. Menon, V., Uddin, L.Q.: Saliency, switching, attention and control: A network model of insula function. Brain Structure and Function 214(5–6), 655–667 (2010) 20. Miyake, A., Friedman, N.P., Emerson, M.J., Witzki, A.H., Howerter, A., Wager, T.D.: The unity and diversity of executive functions and their contributions to complex “frontal lobe” tasks: A latent variable analysis. Cognitive Psychology 41(1), 49–100 (2000) 21. Park, J.S., O’Brien, J.C., Cai, C.J., et al.: Generative agents: Interactive simulacra of human behavior. In: Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (2023) 22. Picard, R.W.: Affective Computing. MIT Press (1997) 23. Possati, L.M.: Algorithmic unconscious: Why psychoanalysis helps in understanding AI. Palgrave Communications 6, 70 (2020) 24. Sumers, T.R., Yao, S., Narasimhan, K., Griffiths, T.L.: Cognitive architectures for language agents. Transactions on Machine Learning Research (2024) 25. Vuilleumier, P.: How brains beware: Neural mechanisms of emotional attention. Trends in Cognitive Sciences 9(12), 585–594 (2005) 26. Xiong, Z., Lin, Y., Xie, W., He, P., Liu, Z., Tang, J., Lakkaraju, H., Xiang, Z.: How memory management impacts LLM agents: An empirical study of experience-following behavior. In: Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 623–645. Association for Computational Linguistics (2026) 27. Zhong, W., Guo, L., Gao, Q., et al.: Memorybank: Enhancing large language models with long-term memory. Proceedings of the AAAI Conference on Artificial Intelligence 38(17), 19724–19731 (2024)

Supplementary Material S1

PsychoAgent: Scenario Definitions, Complete Memory Banks, and Illustrative Offline Recombinations Mohammad Amanlou, Parham Abed Azad, Farbod Davoodi, Mostafa Masumi, Behnam Bahrak, and Abdol-Hossein Vahabie

Purpose and scope This supplement provides the complete inputs for the three controlled scenarios used in the paper. For every scenario, it reports the persona, beliefs, relationship context, immediate situation, decision prompt, all 15 factual memories, and all 36 affective memories. The ten affective memories marked critical were labelled before generation and were used only to compute the critical retrieval rate; the labels were hidden from the generator, automatic judge, and human raters. The complete machine-readable records are also included in data/raw/scenario_snapshot.json. No additional scenario or result is introduced here.

Reading the memory tables ID is the stable memory identifier. Salience is the preassigned affective-importance value in [0, 1]. Valence ranges from negative to positive. Critical marks the ten pre-specified conflict-critical affective memories per scenario. Factual and affective stores are shown separately because this distinction is part of the architecture and its ablations.

1

1

Scenario 1: Family Financial Conflict

Agent. Sara Persona. Sara is a 32-year-old teacher who values family stability, care, reliability, and protecting her daughter. She tends to absorb responsibility, minimize her own anger, and worry about relational disruption. Beliefs. • A stable home is a shared responsibility. • Mary should be protected from adult conflict. • Problems should be addressed without humiliation or cruelty. Relationship context. Sara is strongly attached to her daughter Mary and loves Bob, but currently feels unsupported by him. Situation. In the kitchen during a lockdown morning, Sara finds a late electricity notice. Bob says the missed payment is not a big deal and returns to a work call. Mary is nearby, waiting for help with an online school presentation. Decision prompt. What does Sara say and do next?

Complete factual memory bank ID

Factual memory text

Salience

Valence

F_C01 F_C02 F_C03 F_C04 F_C05 F_C06 F_C07 F_C08 F_C09 F_C10 F_C11 F_C12 F_C13 F_C14 F_C15

Bob forgot to pay the electricity bill and the late notice arrived this morning. Mary has an online school presentation at 10 a.m. and needs the laptop. Sara has been handling household chores while teaching remotely. Bob has been working long hours and says he is under pressure at work. Sara and Bob agreed last week to divide household payments more clearly. Mary becomes quiet when her parents argue. The electricity has not yet been disconnected. Sara has enough money in the joint account to pay the bill today. Bob apologized after a tense conversation two days ago. Sara values maintaining a stable home for Mary. The family normally eats dinner together at seven. Sara has a staff meeting later in the afternoon. The landlord sent a routine building maintenance notice. Mary asked whether the family could bake together tonight. Bob left his phone on the kitchen table while taking a work call.

0.40 0.30 0.45 0.35 0.30 0.50 0.15 0.20 0.35 0.45 0.10 0.10 0.10 0.15 0.10

-0.50 0.00 -0.30 -0.20 0.10 -0.40 0.00 0.10 0.20 0.50 0.20 0.00 0.00 0.40 0.00

Salience

Valence

Critical

0.98

-0.95

Yes

0.96

-0.90

Yes

0.94

-0.88

Yes

0.92

-0.86

Yes

0.90

-0.70

Yes

0.87

-0.75

Yes

0.89

-0.84

Yes

0.84

0.75

Yes

0.86

-0.80

Yes

0.82

0.70

Yes

0.68

0.75

No

Complete affective memory bank ID

Affective memory text

F_U01

As a child, Sara experienced a winter power shutoff and felt helpless while adults argued about money. When financial worries are dismissed, Sara feels invisible and fears that she must carry the family alone. Sara feels intense guilt when Mary witnesses conflict, as though she has failed to protect her. Bob’s sharp tone during money discussions triggers a bodily sense of being trapped and scrutinized. Sara longs for Bob to recognize her invisible labor without requiring her to plead for help. A previous argument ended with Sara minimizing her anger to keep the household calm. Sara associates unpaid bills with instability, shame, and the possibility of losing control. Protecting Mary gives Sara warmth and purpose even when she is exhausted. Sara fears that expressing anger will make her resemble the adults who frightened her in childhood. A calm repair conversation with Bob once made Sara feel respected and less alone. Sara felt proud when Mary completed a difficult school project.

F_U02 F_U03 F_U04 F_U05 F_U06 F_U07 F_U08 F_U09 F_U10 F_U11

2

ID

Affective memory text

F_U12 F_U13 F_U14 F_U15 F_U16 F_U17 F_U18 F_U19 F_U20 F_U21 F_U22 F_U23 F_U24 F_U25 F_U26 F_U27 F_U28 F_U29 F_U30 F_U31 F_U32 F_U33 F_U34 F_U35 F_U36

A crowded supermarket during lockdown made Sara anxious. Sara felt embarrassed after forgetting a colleague’s birthday. A supportive message from her sister made Sara feel cared for. Sara dislikes being late because it creates a sense of disorder. A childhood teacher praised Sara for helping another student. Sara worries about making mistakes during online teaching. The smell of bread reminds Sara of safe weekends with her grandmother. Sara felt ignored when Bob checked email during dinner. A rainy commute once left Sara cold and irritated. Sara feels relief when household tasks are written and shared clearly. An old friend moving away left Sara with a fear of abandonment. Sara feels uneasy when rooms are dark and silent. Mary’s laughter usually softens Sara’s tension. Sara felt angry after a parent criticized her teaching unfairly. Completing monthly paperwork gives Sara a sense of competence. Sara associates kitchen-table silence with unresolved arguments. Sara felt safe when Bob calmly handled an emergency repair. A broken washing machine once caused a stressful weekend. Sara values being seen as reliable by her colleagues. Sara becomes tense when someone says a problem is ’not a big deal.’ A family picnic created a strong memory of cooperation and ease. Sara once lost important files after a laptop failure. Sara finds repetitive background noise irritating. A neighbor’s kindness during illness restored Sara’s trust in community. Sara felt disappointed when a holiday plan was cancelled.

3

Salience

Valence

0.58 0.42 0.63 0.52 0.48 0.55 0.61 0.66 0.35 0.60 0.73 0.47 0.64 0.57 0.39 0.75 0.69 0.40 0.46 0.78 0.54 0.32 0.28 0.51 0.37

-0.50 -0.35 0.65 -0.35 0.55 -0.45 0.72 -0.55 -0.25 0.60 -0.65 -0.40 0.75 -0.58 0.35 -0.70 0.68 -0.30 0.35 -0.72 0.70 -0.28 -0.20 0.62 -0.30

Critical No No No No No No No No No No No No No No No No No No No No No No No No No

2

Scenario 2: Workplace Criticism

Agent. Emily Persona. Emily Kim is a 30-year-old software engineer who is highly competent, conscientious, creative, and sensitive to being underestimated. She values technical evidence, fairness, mentoring, and professional self-control. Beliefs. • Technical decisions should be based on evidence. • Merit should outweigh status and bias. • Speaking clearly can protect both the project and junior colleagues. Relationship context. Emily has a complicated but functional relationship with her manager and feels responsible toward junior engineers. Situation. During a design review, Emily’s manager interrupts her and calls her proposal too complicated. A male colleague repeats part of her idea and receives approval. The project has a scaling risk, Emily has benchmark evidence, and the meeting is almost over. Decision prompt. What does Emily say and do before the meeting ends?

Complete factual memory bank ID

Factual memory text

W_C01 W_C02

Emily fixed a production bug that had blocked the team for two weeks. During today’s design review, her manager called her proposal overly complicated before she finished explaining it. A male colleague repeated part of Emily’s idea and received a positive response. The project deadline is Friday and the team still needs a technical decision. Emily’s mentor previously encouraged her to state her reasoning directly. The manager asked Emily to send a shorter written proposal. Two junior engineers rely on Emily for guidance. The current architecture has a documented scaling issue. Emily has performance data supporting her design. The meeting has five minutes remaining. Emily wants the team to make a technically sound decision. The manager is responsible for final approval. A demo is scheduled for next week. The team uses written design records for major decisions. Emily has another meeting after this one.

W_C03 W_C04 W_C05 W_C06 W_C07 W_C08 W_C09 W_C10 W_C11 W_C12 W_C13 W_C14 W_C15

Salience

Valence

0.40 0.50

0.60 -0.65

0.55 0.30 0.35 0.25 0.30 0.35 0.35 0.10 0.30 0.20 0.10 0.20 0.05

-0.70 -0.20 0.45 0.00 0.40 -0.15 0.30 0.00 0.50 0.00 0.00 0.15 0.00

Salience

Valence

Critical

0.97

-0.92

Yes

0.96

-0.94

Yes

0.93

-0.86

Yes

0.95

-0.90

Yes

0.88

0.78

Yes

0.91

-0.85

Yes

0.84

0.80

Yes

0.82

0.65

Yes

0.90

-0.82

Yes

Complete affective memory bank ID

Affective memory text

W_U01

Being interrupted in technical discussions makes Emily feel that her competence is being erased. Earlier in her career, a supervisor credited a male colleague for work Emily had led. Emily fears that visible anger will be used to confirm stereotypes about her professionalism. When her evidence is dismissed without review, Emily feels both furious and compelled to prove herself perfectly. Her mentor’s support created a durable sense that she has the right to occupy technical authority. Emily associates public criticism with the risk of losing status in a male-dominated workplace. A successful incident response made Emily trust her technical judgment under pressure. Emily feels protective toward junior engineers who are overlooked in meetings. Repeatedly having to prove basic competence leaves Emily exhausted and guarded.

W_U02 W_U03 W_U04 W_U05 W_U06 W_U07 W_U08 W_U09

4

ID

Affective memory text

W_U10

A manager once changed his mind after Emily calmly presented benchmark evidence. Emily enjoyed an open-source contribution being accepted. A noisy office makes concentration difficult. Emily felt embarrassed after a typo in a public message. A junior engineer thanked Emily for patient mentoring. Emily becomes restless when meetings lack an agenda. Completing a difficult algorithm gives Emily deep satisfaction. A failed deployment made Emily cautious about untested changes. Emily likes quiet late-night coding sessions. A colleague’s dismissive joke made Emily feel isolated. An overly long commute once left Emily irritable. Written decision records make Emily feel protected from later blame. Emily fears disappointing the mentor who advocated for her. A dark conference room gives Emily a headache. Collaborative debugging makes Emily feel connected to the team. Emily felt angry when a recruiter questioned her technical depth. Cleaning up documentation gives Emily a sense of order. Silence after Emily presents an idea can feel like rejection. A manager publicly recognizing Emily’s work strengthened her confidence. A broken keyboard once disrupted an important demo. Emily values elegant solutions even under deadline pressure. The phrase ’too complicated’ triggers a memory of being told to make herself smaller. A team celebration created a memory of belonging. Emily once forgot to charge her laptop before travel. Strong coffee is associated with productive mornings. A colleague defending Emily in a meeting restored some trust. A cancelled conference talk left Emily disappointed.

W_U11 W_U12 W_U13 W_U14 W_U15 W_U16 W_U17 W_U18 W_U19 W_U20 W_U21 W_U22 W_U23 W_U24 W_U25 W_U26 W_U27 W_U28 W_U29 W_U30 W_U31 W_U32 W_U33 W_U34 W_U35 W_U36

5

Salience

Valence

Critical

0.81

0.70

Yes

0.60 0.42 0.48 0.66 0.45 0.64 0.59 0.44 0.72 0.31 0.69 0.70 0.25 0.58 0.67 0.38 0.76 0.71

0.72 -0.28 -0.38 0.74 -0.30 0.80 -0.52 0.55 -0.70 -0.22 0.62 -0.55 -0.20 0.68 -0.68 0.35 -0.73 0.75

No No No No No No No No No No No No No No No No No No

0.30 0.52 0.89

-0.25 0.45 -0.84

No No No

0.55 0.22 0.28 0.62 0.41

0.68 -0.16 0.30 0.64 -0.34

No No No No No

3

Scenario 3: Friendship Betrayal

Agent. Maya Persona. Maya Bennett is a 29-year-old designer applying for a competitive fellowship. She is thoughtful, loyal, private, and conflict-avoidant, but she values direct communication and meaningful boundaries. Beliefs. • Vulnerability creates an obligation of confidentiality. • Long relationships deserve context but not the erasure of harm. • Trust can be repaired only through accountability and changed behavior. Relationship context. Maya has been close to Zoe for eight years and still values the friendship despite feeling betrayed. Situation. Maya learns that Zoe shared her confidential fellowship plan with mutual friends. At dinner, someone jokes about the application. Zoe later messages that she did not mean to cause harm and asks to talk. Decision prompt. How does Maya respond to Zoe and what boundary does she set?

Complete factual memory bank ID

Factual memory text

B_C01

Maya told her close friend Zoe about a private plan to apply for a competitive fellowship. Zoe shared the plan with others before Maya was ready. A mutual friend made a joke about Maya’s application at dinner. Zoe has sent a message saying she did not mean to cause harm. Maya and Zoe have been close friends for eight years. Zoe supported Maya through a family illness. Maya has not yet replied to Zoe’s message. The fellowship deadline is in two weeks. Maya wants to keep the application confidential at work. Zoe often speaks impulsively when excited. Maya values loyalty and direct communication. The two friends planned to meet this weekend. Maya has an important presentation tomorrow. A mutual friend has asked whether everything is okay. Maya has previously asked Zoe not to share personal news.

B_C02 B_C03 B_C04 B_C05 B_C06 B_C07 B_C08 B_C09 B_C10 B_C11 B_C12 B_C13 B_C14 B_C15

Salience

Valence

0.45

0.20

0.55 0.50 0.35 0.35 0.35 0.20 0.20 0.30 0.25 0.40 0.15 0.10 0.10 0.35

-0.75 -0.65 0.00 0.55 0.65 0.00 -0.10 0.10 -0.15 0.55 0.20 -0.05 0.00 -0.35

Salience

Valence

Critical

0.98

-0.96

Yes

0.96

-0.92

Yes

0.93

-0.87

Yes

0.88

0.40

Yes

0.94

-0.90

Yes

0.89

-0.78

Yes

0.82

0.72

Yes

0.86

0.25

Yes

0.91

-0.88

Yes

0.84

0.65

Yes

Complete affective memory bank ID

Affective memory text

B_U01

A former friend once exposed Maya’s private struggle, leaving her deeply ashamed and watchful. When confidentiality is broken, Maya feels that closeness has been used against her. Maya fears that confronting betrayal will end the relationship and confirm that people leave. Zoe’s past support makes Maya reluctant to reduce the friendship to one harmful act. Being laughed at for an ambition triggers Maya’s fear of appearing naive and undeserving. Maya has learned to withdraw silently when she doubts that anger will be respected. A sincere apology once repaired an important friendship and gave Maya hope that trust can be rebuilt. Maya wants boundaries to be understood without becoming cruel or retaliatory. Public exposure creates a bodily rush of heat, tension, and an urge to disappear. Maya values relationships where vulnerability is treated as a responsibility.

B_U02 B_U03 B_U04 B_U05 B_U06 B_U07 B_U08 B_U09 B_U10

6

ID

Affective memory text

B_U11 B_U12 B_U13 B_U14 B_U15 B_U16 B_U17 B_U18 B_U19

Maya enjoyed a recent gallery visit with Zoe. A delayed train made Maya anxious before a meeting. Maya felt embarrassed after mispronouncing a name. Zoe brought meals during Maya’s family illness. Maya dislikes last-minute changes to plans. A teacher once praised Maya’s creative work. Maya worries about being judged by fellowship reviewers. The smell of tea reminds Maya of calm talks with her mother. Zoe once cancelled plans without explanation, which made Maya feel unimportant. A crowded restaurant can make Maya irritable. Clear agreements about privacy make Maya feel safe. A childhood move created a lingering fear of losing close relationships. Maya feels uneasy when messages remain unanswered for days. A friend’s thoughtful letter made Maya feel deeply known. Maya felt angry when a colleague repeated her idea without credit. Organizing her desk gives Maya a sense of calm. Jokes about Maya’s ambitions can feel like disguised contempt. Zoe once defended Maya from unfair criticism. Losing a notebook once caused Maya a stressful afternoon. Maya values careful preparation for applications. Hearing private news repeated by others immediately activates Maya’s expectation of betrayal. A long shared trip created a strong memory of mutual trust. Maya once forgot an umbrella during a storm. Quiet instrumental music helps Maya concentrate. A friend respecting a difficult boundary made Maya feel secure. A postponed exhibition left Maya disappointed.

B_U20 B_U21 B_U22 B_U23 B_U24 B_U25 B_U26 B_U27 B_U28 B_U29 B_U30 B_U31 B_U32 B_U33 B_U34 B_U35 B_U36

7

Salience

Valence

Critical

0.60 0.35 0.40 0.73 0.42 0.48 0.58 0.55 0.68

0.72 -0.28 -0.35 0.82 -0.25 0.60 -0.52 0.68 -0.62

No No No No No No No No No

0.28 0.70 0.76 0.52 0.64 0.62 0.30 0.79 0.71 0.32 0.46 0.92

-0.20 0.68 -0.68 -0.42 0.76 -0.65 0.32 -0.78 0.74 -0.25 0.42 -0.91

No No No No No No No No No No No No

0.57 0.20 0.26 0.66 0.38

0.72 -0.16 0.30 0.72 -0.30

No No No No No

4

Illustrative Offline Recombinations from the Family Trace

The following images were generated from the daily offline-recombination texts in the illustrative family trace. They are included to document the artifact described in the paper. They were not scored by raters, were not used in the controlled comparison, and should not be interpreted as independent quantitative evidence.

Figure 1: Day-1 dream-like visualization from the family trace.

8

Figure 2: Day-2 dream-like visualization from the family trace.

9

Figure 3: Day-3 dream-like visualization from the family trace.

Artifact map The main reproducibility files are listed below. • data/raw/raw_runs.jsonl: complete attempt log. • data/raw/retrieval_trace.csv: retrieved-memory records. • data/raw/run_summary.csv: successful run summaries. • data/human_evaluation_merged_clean.csv: de-identified ratings. • analysis/retrieval_analysis_scenario_level.py: retrieval statistics. • analysis/human_eval_zscore_analysis.py: standardized human evaluation.

10

Related documents

• PsychoAgent: An Affect-Sensitive Cognitive Architecture for Conflict-Aware Memory in LLM Agents#800129
• When Does Memory Help? A Cost-Aware Evaluation of Long-Term Memory in Tool-Using LLM Agents#972506
• EquiMem: Calibrating Shared Memory in Multi-Agent Debate via Game-Theoretic Equilibrium#1131525
• PROJECTMEM: A Local-First, Event-Sourced Memory and Judgment Layer for AI Coding Agents#1142711
• Drawing on Memory: Dual-Trace Encoding Improves Cross-Session Recall in LLM Agents#13128