Skip to content

PersonaAgent: Bridging Memory and Action for Personalized LLM Agents

🕒 Published (v1): 2025-06-06 17:29 UTC · Source: Arxiv · Venue: ACL 2026 · link

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

PersonaAgent is a personalized LLM agent framework that couples episodic+semantic memory with a persona (per-user system prompt) that mediates between memory and action. A test-time textual-gradient optimization loop continuously refines each user's persona by minimizing discrepancy between simulated and ground-truth responses. It achieves state-of-the-art on four LaMP personalized decision-making benchmarks without any model fine-tuning.

Problem

Existing LLM agents use a one-size-fits-all policy: general tools, fixed prompts, and no per-user adaptation. Prior personalization approaches (RLHF, user-specific fine-tuning, RAG/PAG workflows) each fail on at least one axis — agentic intelligence, scalability, personal data utilization, or real-time preference alignment — leaving no unified framework that satisfies all four simultaneously.

Method

PersonaAgent operates through three tightly coupled components:

  1. Episodic memory \(D^u = \{(q_i, r_i^{gt}, m_i)\}_{i=1}^{N^u}\): stores timestamped user interactions; retrieves top-\(K\) entries by embedding similarity \(R^u(q^*) = \text{TopK}_{i} \, \text{sim}(h_{q^*}, h_i)\) at inference time.
  2. Semantic memory \(P^u = f_s(S_t, D^u)\): an LLM-generated stable user profile abstracted from episodic history.
  3. Persona \(P\): a per-user system prompt that fuses both memory types and modulates the action policy \(\pi_P(a_t \mid c_t)\) over an augmented action space \(\hat{A} = A \cup D\) (general tools + personalized data access).

Test-time alignment: given a batch \(\{(q_j, \hat{r}_j, r_j^{gt})\}_{j=1}^n\) of recent interactions, the persona is updated iteratively via textual gradient descent: $\(P^* = \arg\min_P \sum_{j=1}^n \mathcal{L}(\hat{r}_j, r_j^{gt} \mid q_j)\)$ where \(\mathcal{L}\) is a textual loss computed by an LLM and gradients are natural-language feedback strings passed to an LLM updater. Optimization runs asynchronously between user sessions so it adds zero online latency.

Key Contributions

  • First end-to-end personalized agent framework unifying episodic memory, semantic memory, personalized action space, and per-user persona in a single architecture.
  • Persona as a dynamic per-user system prompt that enforces personalization at every action-selection step.
  • Test-time user-preference alignment via textual gradient optimization of the persona prompt (no parameter updates to the base LLM).
  • Demonstration of test-time scaling: performance improves with more alignment batch samples, more optimization iterations (up to ~3), and larger retrieved memory \(K\).
  • Model-agnostic gains verified across Mistral-Small, Mistral-Large, Claude-3.5, and Claude-3.7.

Results

  • LaMP-1 (Citation ID): Acc 0.919 / F1 0.918 vs. next-best MemBank 0.862/0.861 and PAG 0.837/0.837.
  • LaMP-2M (Movie Tagging): Acc 0.513 / F1 0.424 vs. MemBank 0.470/0.391 and PAG 0.430/0.387.
  • LaMP-2N (News Categorization): Acc 0.796 / F1 0.532 vs. MemBank 0.741/0.456 and PAG 0.768/0.509.
  • LaMP-3 (Product Rating): MAE 0.241 / RMSE 0.509 vs. MemBank 0.321/0.582 and PAG 0.339/0.835.
  • Ablation: removing the action module causes the largest drops (e.g., F1 on LaMP-2N falls from 0.532 → 0.375); removing alignment drops accuracy on LaMP-1 from 0.919 → 0.894.
  • With Claude-3.7, PersonaAgent reaches 55.0% accuracy on LaMP-2M vs. 32.3% for MemBank.

Limitations

  • Evaluation limited to LaMP classification/rating tasks; no open-ended generation benchmarks in the main results.
  • Textual gradient optimization is an approximation; convergence properties are empirically observed (plateau at ~3 iterations) but not theoretically characterized.
  • Asynchronous alignment assumes session boundaries are well-defined; streaming or continuous-interaction settings are not addressed.
  • Persona stored as free-text system prompt may hit context-length limits as user history grows.
  • Scalability across very large user bases (millions of users each needing individual persona files and episodic buffers) is discussed qualitatively but not benchmarked.

Relevance to Harnesses / Meta-Harnesses

PersonaAgent is a concrete instance of a meta-harness pattern: a controller layer (the persona + alignment loop) that wraps an underlying agent runtime and reconfigures it per-user without touching the base model. The test-time alignment loop — simulate → compute textual loss → update system prompt → repeat — is structurally identical to a meta-harness's outer optimization loop that tunes the prompt/config governing an inner agent. For researchers tracking harness architectures, the persona-as-system-prompt abstraction offers a clean model for how a meta-level controller can impose user-specific policies over a fixed action-tool substrate, and the textual gradient mechanism demonstrates a gradient-free prompt-optimization loop that could be generalized to tune other harness-level configurations (tool selection policies, memory retrieval thresholds) at test time.