Skip to content

ClinicalAgents: Multi-Agent Orchestration for Clinical Decision Making with Dual-Memory

🕒 Published (v1): 2026-03-27 08:52 UTC · Source: Arxiv · Venue: KDD 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

ClinicalAgents is a multi-agent framework for clinical decision-making that replaces rigid sequential workflows with a Monte Carlo Tree Search (MCTS)-based orchestrator, enabling hypothesis-driven, evidence-seeking reasoning with backtracking. It combines a mutable working memory (patient state) and a static experience memory (guidelines + historical cases) to reduce information fragmentation across multi-agent turns. It achieves state-of-the-art performance on a multi-stage clinical workflow benchmark.

Problem

Existing MAS for clinical decision-making (CDM) suffer from three compounding failures: (1) pre-scripted, linear interaction protocols that cannot replan or backtrack when new evidence emerges; (2) window-bounded dialogue context that fragments critical details across multi-turn agent interactions; (3) static generation framing that misses the iterative hypothetico-deductive nature of real clinical diagnosis, causing premature closure on insufficient evidence.

Method

ClinicalAgents orchestrates a pool of specialized agents (referral, imaging, diagnosis, treatment) through three components:

Clinical Orchestrator (MCTS-based): Formulates orchestration as an MDP \(\langle M^t_\text{work}, \mathcal{A}, R, T, \eta \rangle\). The action space \(\mathcal{A} = \mathcal{A}_\text{agent} \cup \{a_\text{rag}, a_\text{back}, a_\text{term}\}\) includes agent calls, knowledge retrieval, backtracking, and termination. The orchestrator selects top-\(K\) candidate actions, runs \(N\) rollouts per candidate to estimate \(Q(M^t_\text{work}, a)\) as discounted cumulative reward, and uses LLM-prior-guided PUCT for node selection: $\(a^* = \arg\max_{a \in \mathcal{A}_K} \left[ Q(M^t_\text{work}, a) + \lambda \cdot \mathcal{O}(a \mid M^t_\text{work}, I_\text{select}) \right]\)$ The reward function \(R_t = \alpha \cdot \max(0, \Delta E^m_t) + \beta \cdot \max(0, \Delta c_t) - \gamma \cdot \mathbf{1}[\Delta E^m_t \leq 0 \wedge \Delta c_t \leq 0]\) rewards reductions in missing evidence \(\Delta E^m_t\) and gains in top-hypothesis confidence \(\Delta c_t\). Backtracking is triggered when \(|E^m_t| > 0\) and routes to the appropriate prior stage based on the type of missing evidence.

Dual-Memory System: - Working memory \(M^t_\text{work} = \langle E^t, H^t, \tau^t \rangle\): mutable; accumulates evidence, hypotheses, and the action trajectory across all agent interactions. - Experience memory \(M_\text{exp} = \langle D_\text{guide}, D_\text{CDC} \rangle\): static; retrieves guideline chunks via AGRAG (a knowledge-graph-augmented RAG) and identifies potentially missing evidence \(E^p_t\) by scoring candidate evidence items from similar historical cases using cosine-weighted importance: \(\text{Imp}(e \mid E^t, \tilde{D}_\text{cdc}, H^t) = \frac{1}{n_\text{cdc}} \sum_i \text{sim}(E^t \oplus H^t, d_i) \cdot \mathbf{1}[e \in E^*_{d_i}] \cdot \mathbf{1}[e \notin E^t]\).

The reasoning cycle is: Perceive → Hypothesize → Verify → Update, with backtracking re-entering Perceive when evidence is insufficient.

Key Contributions

  • MCTS-based dynamic orchestrator that replaces pre-scripted MAS workflows with a search-theoretic policy over agent actions, enabling evidence-triggered backtracking and replanning.
  • Reward function explicitly encoding the clinical goal of narrowing differential diagnoses via missing-evidence reduction and confidence gain.
  • Dual-memory architecture separating short-term mutable patient state (working memory) from long-term reusable knowledge (experience memory), reducing cross-turn context fragmentation.
  • Hybrid RAG mechanism combining passive guideline augmentation (AGRAG) with active evidence-seeking from historical cases (CDC importance scoring).
  • State-of-the-art results on a multi-stage clinical workflow benchmark over single-agent and multi-agent baselines.

Results

The paper text provided is truncated before the results tables; no specific benchmark scores, dataset names, or numerical comparisons against baselines are available in the supplied excerpt. The abstract and contributions claim "best performance among evaluated baselines" and significant gains in diagnostic accuracy and explainability over strong single-agent and multi-agent baselines, but no concrete numbers can be faithfully reported from the provided text.

Limitations

  • Experimental details and numbers are not available in the provided text excerpt, preventing independent verification of claimed gains.
  • MCTS with \(N\) rollouts per candidate action is computationally expensive; cost relative to sequential MAS baselines is not discussed in the provided text.
  • Experience memory \(M_\text{exp}\) is static — it does not update with new cases encountered at inference time, limiting adaptation to rare or novel diseases.
  • The backtracking mechanism relies on an LLM to detect missing evidence (\(E^m_t\)) and assign importance scores, introducing potential hallucination in the very step designed to reduce diagnostic error.
  • Evaluation is on a single benchmark dataset (details not visible in excerpt); generalizability across clinical specialties or EHR systems is unclear.

Relevance to Harnesses / Meta-Harnesses

ClinicalAgents is a direct existence proof that replacing a static DAG-style harness with a search-based orchestrator (MCTS) unlocks qualitatively different agent behavior — specifically, hypothesis generation, targeted evidence retrieval, and principled backtracking — none of which are achievable in fixed-pipeline harness designs. The dual-memory split (mutable per-episode working memory vs. static experience memory with guidelines and cases) is a concrete architectural pattern for any multi-agent harness facing state fragmentation across turns: it cleanly separates "what we know so far" from "what we always know," which scales better than passing a growing context blob between agents. For someone building agent harnesses, the key transferable idea is the orchestrator-as-search-policy framing — treating orchestration as a tree search over agent-action sequences rather than a predetermined call graph — which generalizes well beyond clinical domains to any harness where the right agent sequence is input-dependent and not knowable in advance.