PersonaAgent Framework

• PersonaAgent is a personalized LLM framework that integrates individualized memory, dynamic prompt engineering, and real-time preference alignment to overcome generic model limitations.
• It features a modular architecture with dual memory systems and a GraphRAG extension for effective knowledge retrieval and community-aware summarization.
• Empirical results show significant performance gains on personalization benchmarks, making it ideal for applications in education, research, and content recommendation.

PersonaAgent is a personalized LLM agent framework that integrates individualized memory, flexible prompt engineering, and real-time preference alignment to address the limitations of generic LLMs in user-specific settings. Variants such as PersonaAgent with GraphRAG further augment the base design by incorporating knowledge-graph-based retrieval and community-aware summarization, yielding state-of-the-art results on personalization benchmarks (Liang et al., 21 Nov 2025, Zhang et al., 6 Jun 2025).

1. System Architecture

PersonaAgent consists of modular components tightly coupled to provide user-adaptive behavior:

• LLM Core: A pre-trained or fine-tuned LLM (e.g., LLaMA, Claude) that produces responses given persona-aware prompts.
• Personalized Memory Module: Separates into episodic memory ($\mathcal{D}^u$) for time-stamped, fine-grained user interactions, and semantic memory ($\mathcal{P}^u$) for high-level user profiles synthesized via summarization.
• Persona (System Prompt $P$): Dynamically constructed per user, embedding both episodic and semantic memory, and subject to real-time updates.
• Personalized Action Module: Maintains an action policy ($\pi_P$) over an augmented space including tool actions (e.g., memory RAG API, Wiki search) and observes interaction outcomes to update state.
• GraphRAG Indexer and Community Detector: (GraphRAG variant) Maintains a heterogeneous knowledge graph $G=(V,E)$ from user and global interactions, employs dense retrieval, and uses graph clustering (e.g., modularity maximization) to detect communities.
• Prompt Generator: Assembles user history, community, and global context into the system prompt.

High-level data flow: On a user query, memory modules and/or the GraphRAG indexer retrieve pertinent past and global contexts, which are merged into a persona prompt. The LLM core, conditioned on this prompt and current query, generates actions and the final response. Executed tool calls and new observations update episodic memory or the knowledge graph (Liang et al., 21 Nov 2025, Zhang et al., 6 Jun 2025).

2. Knowledge Graphs and Memory

For the GraphRAG extension, the knowledge graph $G=(V,E)$ includes:

• Nodes:
  • Interaction nodes ($v_i$): attributes $t_i$ (text), $s_i$ (timestamp), $c_i$ (category label).
  • Concept nodes ($\mathcal{P}^u$0): entities or keywords extracted from $\mathcal{P}^u$1.
  • Category nodes ($\mathcal{P}^u$2): high-level domains (e.g., "politics", "action-movies").
• Edges:
  • Interaction–Category: $\mathcal{P}^u$3 if $\mathcal{P}^u$4.
  • Interaction–Concept: $\mathcal{P}^u$5 when $\mathcal{P}^u$6 appears in $\mathcal{P}^u$7.
  • Concept–Concept: $\mathcal{P}^u$8 if concepts $\mathcal{P}^u$9 co-occur or share a category.
• Embeddings and Edge Weights:
  • Nodes represented as vectors $P$0. For interactions, $P$1 or dense encoding.
  • Cosine similarity: $P$2.
  • Edge weights: For concept–concept, Jaccard index $P$3; for others, uniform or similarity-based.

Memory mechanisms in base PersonaAgent mirror this structure: episodic memory for granular interactions and semantic memory for aggregate profiles, both used in prompt construction (Liang et al., 21 Nov 2025, Zhang et al., 6 Jun 2025).

3. Retrieval-Augmented Generation and Community Awareness

GraphRAG fuses individual and collective intelligence by interleaving graph-based retrieval and clustering:

1. Personal Retrieval: Given user $P$4 and query $P$5, retrieve top-K relevant personal interactions by similarity:

$P$6

2. Global Retrieval: Retrieve top-K relevant non-user interactions:

$P$7

1. Subgraph Extraction and Expansion: Form $P$8 by combining seed nodes $P$9 and their neighbors.
2. Community Detection: Use modularity maximization (e.g., Louvain) to cluster $\pi_P$0, optimizing

$\pi_P$1

where $\pi_P$2 is the degree, $\pi_P$3 the total edge weight, and $\pi_P$4 indicates community membership.

1. Summarization: LLM generates a summary $\pi_P$5 for each community $\pi_P$6, maximizing conditional log-likelihood:

$\pi_P$7

This process enables the agent to identify not only individual habits but also salient collective patterns relevant to the user's context (Liang et al., 21 Nov 2025).

4. Persona-Aligned Prompt Engineering

PersonaAgent prompt engineering synthesizes five elements into the system prompt $\pi_P$8:

1. Task Instructions: Precise guidance for the LLM (e.g., "Classify the movie into tags...").
2. User Profile Summary: Weighted aggregation of user history, such as $\pi_P$9.
3. Recent Personal Interactions: A recency-weighted, relevance-scored list (“Recent: 1. [2024-05-01] ‘...’ (score 0.83)...”).
4. Community Insights: For each cluster $G=(V,E)$0, a summary produced by the LLM (“Cluster $G=(V,E)$1:” followed by $G=(V,E)$2).
5. Global Preferences: Aggregated interaction statistics across the population (“Common across users: politics (0.42), youth activism (0.33)”).

A single prompt template integrates these fields, ensuring that the LLM response is grounded in both individual and collective knowledge (Liang et al., 21 Nov 2025).

Prompt refinement during inference (test-time preference alignment) is implemented through gradient-based updates: simulating the last $G=(V,E)$3 interactions, measuring textual loss ($G=(V,E)$4) versus ground truth, and updating $G=(V,E)$5 via

$G=(V,E)$6

This process personalizes agent outputs in real time without any modification of the underlying LLM parameters (Zhang et al., 6 Jun 2025).

5. Evaluation and Empirical Results

PersonaAgent and its GraphRAG variant have been evaluated on the LaMP benchmark suite, including tasks for news categorization, movie tagging, and product rating:

Task	Base F1/MAE	+GraphRAG F1/MAE	Δ F1/MAE	% Gain/Reduction
News (LaMP-2N)	0.532	0.591	+0.059	11.1%
Movies (LaMP-2M)	0.424	0.662	+0.238	56.1%
Product (LaMP-3)	0.241 (MAE)	0.216 (MAE)	-0.025	10.4%

PersonaAgent consistently outperforms non-personalized, prompt-based, and previous agentic baselines, including ICL, ReAct, MemBank, and earlier RAG workflows. Removal of any module (test-time alignment, persona, memory, action) leads to significant degradation in performance, confirming the necessity of each component (Liang et al., 21 Nov 2025, Zhang et al., 6 Jun 2025). Gains persist across multiple LLM backbones and scale with increased alignment batch size and retrieval depth.

Qualitative studies demonstrate that integrating community insights corrects for biased personal histories (e.g., relabeling content once broader patterns are considered) (Liang et al., 21 Nov 2025).

6. Applications, Scalability, and Limitations

Domain applications of PersonaAgent include personalized education, research assistance (e.g., citation recommendations), content recommendation (movies, news, products), and professional productivity domains (legal, medical) (Zhang et al., 6 Jun 2025).

Scalability is achieved through:

• Prompt-only personalization (no per-user LLM fine-tuning).
• Lightweight gradient-based test-time updates (batch size $G=(V,E)$7, 1–3 iterations).
• Controlled memory growth via top-K retrieval and summarization.

Privacy and future considerations: The system currently relies on text-only feedback for preference alignment. Storing personal histories introduces privacy risks, which could be mitigated using federated or differentially private retrieval processes. Expanding multimodal support is an open research direction (Zhang et al., 6 Jun 2025).

This suggests that PersonaAgent and its extensions provide a highly modular, effective framework for scalable, real-world personalization without the computational overhead of user-specific LLM retraining. The community-aware GraphRAG mechanism further demonstrates that structured global knowledge, in combination with individual user history, is critical for robust and socially aligned personalized AI agents.