DualMind: Dual-Process Cognitive Architectures

Updated 5 July 2026

DualMind is a family of dual-process models that decompose cognition into complementary fast and slow components to improve task performance.
The approach integrates lightweight recall with deeper deliberative processing, achieving higher fidelity in memory, reasoning, and real-time action across multiple domains.
Wide applications include LLM memory systems, conversational agents, and public opinion simulations, with empirical gains noted in metrics like cross-session inference and response latency.

“DualMind” denotes a cluster of cognitively inspired architectures in which intelligence is decomposed into two coupled processes, perspectives, or phases rather than treated as a monolithic inference pipeline. In current literature, the term appears both as the exact name of a multi-agent public-opinion simulation platform and as a broader design pattern spanning LLM memory systems, conversational agents, emotional-support dialogue, spoken reasoning, and generalist control. Across these uses, the recurring premise is that a single retrieval surface, a single response mode, or a single training stage is insufficient for tasks involving long-horizon revision, context-sensitive reasoning, interpersonal modeling, or real-time action generation (Huang et al., 28 Jan 2026, Fei et al., 8 Jun 2026, You et al., 19 Mar 2026).

1. Conceptual scope and recurring meanings

Across recent work, “DualMind” and closely related “dual-process,” “dual-mind,” and “dual-brain” formulations do not refer to one standardized architecture. Instead, they instantiate a shared structural motif: one subsystem handles fast, local, or routine processing, while another handles slower, higher-fidelity, more abstract, or complementary processing.

Work	Dual structure	Domain
DCPM (Fei et al., 8 Jun 2026)	System 1 / System 2	long-term memory for LLM agents
D-Mem (You et al., 19 Mar 2026)	Mem0 $^\ast$ / Full Deliberation with Quality Gating	long-horizon reasoning
Mind2 (Hong et al., 17 Mar 2025)	User $\rightarrow$ System / System $\rightarrow$ User perspectives	emotional support dialogue
DUMA (Tian et al., 2023)	Fast Mind / Slow Mind	Chinese real-estate online conversation
DualMind (Huang et al., 28 Jan 2026)	persistent cognitive beliefs / transient affective responses	public-opinion dissemination
DualMind (Wei et al., 2023)	Phase I / Phase II	generalist decision-making
MPS (Wu et al., 10 Oct 2025)	Formulation Brain / Articulation Brain	real-time spoken reasoning

These systems differ in what is being split. In DCPM and D-Mem, the split is primarily between lightweight online recall and heavier consolidation or deliberation (Fei et al., 8 Jun 2026, You et al., 19 Mar 2026). In Mind2, the split is interpersonal and bidirectional, emphasizing mutual mind-modeling rather than latency control (Hong et al., 17 Mar 2025). In the public-opinion platform "DualMind," the duality is between cognitive beliefs and affective responses inside each agent state (Huang et al., 28 Jan 2026). In the generalist agent also named "DualMind," the duality is temporal and pedagogical: common control knowledge is learned first, and prompt-conditioned behavior is learned second (Wei et al., 2023).

This suggests that “DualMind” functions less as a single framework name than as a family resemblance term for systems that reject one-pass, single-mode cognition.

2. Dual-process memory and deliberative agent architectures

In long-term memory for LLM agents, the strongest “DualMind”-style formulation is DCPM, or Dual-Process Cognitive Memory. DCPM treats memory as a cognitive system rather than a retrieval index and organizes it as a capability hierarchy ascending from raw inputs and atomic facts through diachronic belief trajectories and identity to schemas, latent intentions, and cross-domain core schemas (Fei et al., 8 Jun 2026). Its System 1 is a synchronous daytime writer invoked on explicit add_memory requests. System 1 performs raw persistence, extraction with an LLM $M_e$ , reconciliation with a reconcile LLM $M_r$ , and batch upsert with revision pointers. The core formal device is the bidirectional supersedes link,

$n_{\text{new}.Supersedes = n_{\text{old}}, \quad n_{\text{old}.SupersededBy = n_{\text{new}}$

which creates a doubly linked list of revisions over an otherwise flat vector store. System 2 is an asynchronous nighttime engine operating offline over accumulated outputs of System 1. It preprocesses fresh facts with two-stage DBSCAN, induces schemas and intentions with an induction agent $M_a$ , and then performs a cross-domain sweeper that induces core schemas when behavioral similarity is high but surface semantic similarity is low. On LongMemEval, PersonaMem, and PersonaMem-v2, enabling System 2 helps most on implicit cross-session inference, with deltas of $+0.44 / +0.33$ , $+1.31 / +2.41$ , and $+2.51 / +5.20$ , respectively; on PersonaMem-v2, removing the cross-domain collision/core-schema mechanism costs about 4.1 points, while removing supersedes chains costs about 2.9 points (Fei et al., 8 Jun 2026).

D-Mem adopts a different dual-process decomposition. Its System 1 is Mem0 $\rightarrow$ 0, a lightweight retrieval-based memory module; its System 2 is Full Deliberation, a high-fidelity fallback that reads raw dialogue history chunk-by-chunk; and the two are bridged by a Multi-dimensional Quality Gating policy (You et al., 19 Mar 2026). The gating function

$\rightarrow$ 1

accepts the initial answer when relevance, faithfulness/consistency, and completeness all pass, and otherwise escalates to Full Deliberation. Full Deliberation splits history into chunks of 60 messages, uses a sliding context window of 4 preceding messages, extracts question-relevant facts scored from 0 to 10, keeps facts with score $\rightarrow$ 2, and then filters further with a threshold $\rightarrow$ 3. On LoCoMo with GPT-4o-mini, Quality Gating achieves F1 53.5, outperforming Mem0 $\rightarrow$ 4 at 51.2 and recovering 96.7\% of Full Deliberation’s 55.3, while using 12,681 tokens versus 35,435 for Full Deliberation, roughly a 64\% reduction (You et al., 19 Mar 2026).

DUMA, a Dual-Mind Conversational Agent with Fast and Slow Thinking, extends the same logic into a two-LLM conversational stack. Mind_Fast is the primary user-facing model, while Mind_Slow is an internal deliberative model invoked by a boolean Invoke signal emitted by the fast model itself (Tian et al., 2023). Slow Mind follows a ReAct-style loop of Reason, Act, Obs, and Finish, can use tools, and writes its outputs into Dialogue Memory for later reuse. In Chinese real-estate online conversation, DUMA is evaluated against ChatGPT $\rightarrow$ 5 on House Expertise, Tool Calling Ability, Industry Familiarity, Service Attitude, Demand Mining, and Promote Invitation. Reported scores show DUMA outperforming ChatGPT $\rightarrow$ 6 across all six metrics, with the full two-stage version achieving 1.550, 1.417, 1.125, 1.810, 1.357, and 1.471, respectively (Tian et al., 2023).

Taken together, these systems converge on one criticism of retrieval-only or single-pass agents: query-agnostic compression and flat recall are insufficient for knowledge update, temporal reasoning, or difficult multi-hop reconstruction.

3. Bidirectional mental-state modeling in dialogue

Mind2 defines a different but related sense of “DualMind.” Rather than fast versus slow cognition, it models dialogue as an interaction between two minds, each reasoning about the other (Hong et al., 17 Mar 2025). The framework is designed for emotional support dialogue and performs bidirectional cognitive discourse analysis from both the User $\rightarrow$ 7 System and System $\rightarrow$ 8 User perspectives.

Its input sequence augments dialogue history with speaker-role markers and situational context,

$\rightarrow$ 9

then appends cognitive knowledge,

$\rightarrow$ 0

Generation is performed by a BlenderBot-based Seq2Seq backbone with standard negative log-likelihood training,

$\rightarrow$ 1

A central mechanism is the dynamic discourse context propagation window $\rightarrow$ 2,

$\rightarrow$ 3

which provides a bounded moving local context around the current utterance. Cognitive extraction over $\rightarrow$ 4 yields Bidirectional Cognitive Knowledge with three components: Bidirectional Theory-of-Mind, BTM-based Psychological Expected Utility, and BTM-based Cognitive Rationality. The extracted terms remain traceable to the original discourse window, which the paper treats as an interpretability advantage. Evaluated on ESConv, a dataset of 1,300 crowd-sourced emotional support conversations, Mind2 is trained with only 10\% of the original training data while baselines are reported using 100\% of training data; the paper reports competitive performance versus state-of-the-art emotional-support systems, with ablations indicating that removing BTM tends to hurt more than removing PEU or BCR (Hong et al., 17 Mar 2025).

Within the broader DualMind motif, Mind2 shows that the “dual” need not be computational speed versus deliberation. It can also denote explicit mutual mind-reading and speaker-indexed cognitive representations.

4. DualMind in control and spoken reasoning

The 2023 generalist agent named DualMind applies duality at the training level rather than at inference-time routing (Wei et al., 2023). The model is built around an Encoder-Decoder Control Transformer with a ViT-B/16 state tokenizer, TokenLearner compression from 196 ViT patch tokens to 8 state tokens, action discretization into 256 uniform bins, pretrained CLIP prompt encoding, and cross-attention prompt fusion. Phase I trains the entire model with a self-supervised objective

$\rightarrow$ 5

to learn generic information of state-action transitions. Phase II learns a prompt-conditioned policy using

$\rightarrow$ 6

while retraining only a small fraction of the model. Evaluated on MetaWorld and Habitat with a total of 100k episodes, DualMind is reported to outperform previous generalist agents by over 50\% on Habitat and over 70\% on MetaWorld; on the 45 tasks in MetaWorld, it achieves over 30 tasks at a 90\% success rate, with 39 tasks above 0.5 SR, 34 above 0.8 SR, and 28 above 1 SR (Wei et al., 2023).

Mind-Paced Speaking extends the pattern to real-time spoken LLMs. Its “dual-brain” design separates a Formulation Brain, which generates the Chain-of-Thought stream, from an Articulation Brain, which incrementally generates spoken response segments conditioned on partial reasoning (Wu et al., 10 Oct 2025). The CoT is segmented as

$\rightarrow$ 7

and response generation is paced segment-by-segment. The articulation side is trained with think-incomplete supervised fine-tuning so that it can operate on partial thought rather than full rationales. Using Step-Audio 2 as backbone, the paper reports on Spoken-MQA that MPS-tbs achieves 93.0 average accuracy, MPS-thkfirst 93.9, and MPS-spkfirst 92.8, compared with 68.6 for Mini-Omni-Reasoner and 70.6 for MPS-wo/thk. The extra tokens generated before the first response token are 762 for MPS-tbs, 80 for MPS-thkfirst, and 0 for MPS-spkfirst (Wu et al., 10 Oct 2025).

These two systems illustrate that “DualMind” can also mean decomposing competence acquisition into two phases or decomposing reasoning and articulation into parallel streams.

5. DualMind as a public-opinion simulation platform

The paper explicitly titled "DualMind: Towards Understanding Cognitive-Affective Cascades in Public Opinion Dissemination via Multi-Agent Simulation" uses the term in a narrower and domain-specific sense (Huang et al., 28 Jan 2026). Here the duality is not fast versus slow reasoning but the interaction between persistent cognitive beliefs and transient affective responses in public-relations crises.

Each agent $\rightarrow$ 8 at time $\rightarrow$ 9 maintains a semantic/cognitive state $M_e$ 0, an emotional state $M_e$ 1, and episodic memory. Retrieved context is computed by semantic matching plus recency decay: $M_e$ 2 where the normalized weights depend on $M_e$ 3, semantic similarity, and exponential recency decay $M_e$ 4. Belief and affect are then updated jointly through an affect-gated rule: $M_e$ 5 followed by normalization of $M_e$ 6 onto the unit sphere. Propagation is governed by the Polarized Affective Cascade Model,

$M_e$ 7

and macro-level cascade explosiveness is summarized by a reproduction coefficient

$M_e$ 8

The platform is evaluated on 15 real-world crises occurring after August 2024, evenly split across the United States, China, and Europe, using social media as ground truth. The protocol uses 100 agents per crisis, averages over 5 stochastic seeds, and compares against LAID, LPOD, and LLM-GA. Reported averages are overall trajectory similarity $M_e$ 9 and overall outcome divergence $M_r$ 0, with DualMind said to outperform all baselines on all 15 cases (Huang et al., 28 Jan 2026).

This is the most literal instance of “DualMind” as a named system. It also broadens the concept beyond LLM internal architecture into a formalized social simulation of crisis opinion dynamics.

6. Shared design principles, empirical patterns, and limitations

Despite heterogeneity of domain, several regularities recur. First, these works consistently target failures attributed to single-surface processing: DCPM criticizes collapsing belief revision, causal coupling, and cross-domain abstraction into retrieval tuned for surface recall; D-Mem critiques query-agnostic compression and lossy abstraction; Mind2 rejects one-way user-state estimation; and the public-opinion DualMind argues that a single scalar sentiment misses the interaction of affect and belief (Fei et al., 8 Jun 2026, You et al., 19 Mar 2026, Hong et al., 17 Mar 2025, Huang et al., 28 Jan 2026).

Second, the largest empirical gains usually appear on tasks that reward implicit structure rather than direct recall. DCPM reports that enabling System 2 helps least on span recall and most on implicit cross-session inference, reaching up to $M_r$ 1 on PersonaMem-v2 (Fei et al., 8 Jun 2026). D-Mem shows that Quality Gating occupies the best part of the accuracy-cost curve by approaching Full Deliberation while avoiding its over-10 $M_r$ 2 token and latency cost (You et al., 19 Mar 2026). MPS reports near-think-before-speaking reasoning quality while drastically reducing pre-response latency (Wu et al., 10 Oct 2025). This suggests that dual-process designs are most valuable when the bottleneck is not storage capacity alone but the timing, fidelity, or abstraction level of computation.

Third, the “dual” relation varies substantially across papers. Some systems use explicit routing, as in D-Mem’s gating function and DUMA’s boolean Invoke signal (You et al., 19 Mar 2026, Tian et al., 2023). Others use asynchronous consolidation, as in DCPM’s nighttime engine (Fei et al., 8 Jun 2026). Still others formalize duality as speaker-indexed cognition, as in Mind2, or as affect-gated belief evolution, as in the public-opinion platform (Hong et al., 17 Mar 2025, Huang et al., 28 Jan 2026). A plausible implication is that “DualMind” names an architectural commitment to differentiated cognitive roles rather than a single canonical algorithm.

The literature also records important limitations. DUMA provides no explicit numeric threshold or formal complexity score for invoking Slow Mind (Tian et al., 2023). The public-opinion DualMind excerpt does not report a formal ablation study and explicitly notes simplified cognition, abstracted platform algorithms, and network scale limitations (Huang et al., 28 Jan 2026). In MPS, early response segments may be lower quality because they are based on incomplete reasoning (Wu et al., 10 Oct 2025). In DCPM, System 2 contributes least on straightforward span recall, which matches the architectural prediction but also delimits where abstraction is useful (Fei et al., 8 Jun 2026).

The term “DualMind” therefore denotes a research tendency rather than a single settled doctrine: the systematic decomposition of intelligent behavior into two interacting modes, stores, perspectives, or learning stages, with the aim of improving fidelity on tasks that exceed what flat retrieval, single-pass generation, or undifferentiated state modeling can support.