Memory Papers
Memory Architecture

PersMem: Internalizing Personality into Dual-Pathway Memory for LLM Agents

AuthorsHanzhong Zhang, Ziwei Xiang, Weicheng Xie et al.

arXiv 20262026
arXivPDF

TL;DR

PersMem uses a personality-conditioned dual-pathway memory pipeline with affective appraisal, retention, passive and active retrieval to reach 48.1% attachment-profile classification and 67.5% Big Five dialogue accuracy.

SharePost on XLinkedIn

Read our summary here, or open the publisher PDF on the next tab.

THE PROBLEM

Personas in prompts do not control memory processing, causing inconsistent behaviour

Existing role-playing agents keep memory retrieval largely independent of the pre-defined personality, so memory processing often contradicts the assigned traits.

This breaks long-horizon role-playing and emotional companionship, because agents retrieve neutral or positive memories even when a negative personality-consistent response is expected.

HOW IT WORKS

PersMem — Personality-Integrated Dual-Pathway Memory

PersMem encodes a fixed trait vector and feeds it into Affective Appraisal, Memory Retention, Passive Affect-Driven Memory Retrieval, and Active Goal-Driven Memory Retrieval before response generation.

You can think of PersMem as a personality-tuned hippocampus, where traits bias which memories are stored, forgotten, and resurfaced, instead of just tinting the final reply.

This personality-integrated pipeline lets PersMem expose inspectable memory traces and produce personality-consistent recall patterns that a plain context window or semantic RAG cannot.

DIAGRAM

Turn-Level Interaction and Dual-Pathway Retrieval Flow

This diagram shows how PersMem processes a single user turn, from affective appraisal through passive and active retrieval to response and storage.

DIAGRAM

Evaluation Pipeline Across Attachment, Big Five, and CoSER

This diagram shows how PersMem is evaluated on attachment memory processing, Big Five dialogues, and CoSER role-playing.

PROCESS

How PersMem Handles a Dialogue Turn

  1. 01

    Affective Appraisal

    PersMem converts the user input into a VAD vector and adjusts it using the personality vector inside Affective Appraisal, updating the current affective state.

  2. 02

    Memory Retention

    PersMem applies Memory Retention to all stored memories, using personality-dependent decay rates to keep or forget events based on arousal and trait settings.

  3. 03

    Passive Affect-Driven Memory Retrieval

    PersMem runs Passive Affect-Driven Memory Retrieval, combining semantic similarity, retention, and affective relation scores to sample an initial memory set.

  4. 04

    Active Goal-Driven Memory Retrieval

    PersMem launches Active Goal-Driven Memory Retrieval, where a personality-conditioned redirection gate can steer search using passively retrieved memories before sending context to the response model.

KEY CONTRIBUTIONS

Key Contributions

  • 01

    Personality-Integrated Memory Architecture

    PersMem represents personality as a fixed low-dimensional trait vector and maps it into Affective Appraisal, Memory Retention, Passive Affect-Driven Memory Retrieval, and Active Goal-Driven Memory Retrieval without exposing traits to the response model.

  • 02

    Personality-Conditioned Redirection Mechanism

    PersMem introduces a redirection gate in Active Goal-Driven Memory Retrieval whose firing threshold depends on attachment anxiety or neuroticism, enabling trait-dependent steering of multi-step recall.

  • 03

    Personality-Dependent Evaluation of Memory Processing

    PersMem is evaluated on attachment and Big Five settings, achieving 48.1% four-way attachment classification and 67.5% Big Five dialogue accuracy, plus 66.13 average on CoSER role-playing.

RESULTS

By the Numbers

Attachment classification accuracy

48.1%

+23.1 percentage points over chance baseline

Big Five dialogue accuracy

67.5%

+6.7 percentage points over random-memory control

CoSER average score

66.13

+6.18 over CoSER GPT-4o average 59.95

CoSER Character Fidelity

69.33

+20.38 over CoSER GPT-4o Character Fidelity 48.95

On attachment memory-processing classification, PersMem reaches 48.1% accuracy versus 25.0% chance, showing distinct trait-conditioned recall patterns. On Big Five dialogues and CoSER, PersMem’s 67.5% dialogue accuracy and 66.13 average CoSER score demonstrate that personality-integrated memory processing yields detectable personality expression and strong role-playing quality.

BENCHMARK

By the Numbers

On attachment memory-processing classification, PersMem reaches 48.1% accuracy versus 25.0% chance, showing distinct trait-conditioned recall patterns. On Big Five dialogues and CoSER, PersMem’s 67.5% dialogue accuracy and 66.13 average CoSER score demonstrate that personality-integrated memory processing yields detectable personality expression and strong role-playing quality.

BENCHMARK

Attachment Profile Classification on Held-Out Test Set

Accuracy (%) for predicting attachment profile from PersMem memory traces.

BENCHMARK

CoSER Role-Playing Evaluation Scores

Scores on CoSER dimensions for PersMem versus reference CoSER GPT-4o.

KEY INSIGHT

The Counterintuitive Finding

PersMem achieves 48.1% four-way attachment classification using only memory-processing traces, while a classifier using recalled-memory text embeddings alone stays at 25.0% chance.

This is surprising because many assume personality is mainly visible in language style, but PersMem shows that internal memory operations carry stronger personality signals than surface text.

WHY IT MATTERS

What this unlocks for the field

PersMem unlocks agents whose memory storage, forgetting, and recall are all explicitly shaped by a stable personality vector, not just prompt wording.

Builders can now debug and design personality-consistent agents by inspecting affective appraisal, retention, and retrieval traces, enabling controllable long-term personas and research on personality-conditioned memory dynamics.

~14 min read← Back to papers

Related papers

Memory Architecture

A Control Architecture for Training-Free Memory Use

Yanzhen Lu, Muchen Jiang et al.

· 2026

TAG routes low-confidence steps to uncertainty-based routing, filters them with guarded acceptance with rollback, chooses between bank selection across rule and exemplar memory, and prunes via evidence-based retirement inside a unified control loop. On SVAMP and ASDiv, TAG reaches 81.0% and 85.2% accuracy, improving over the 74.0% and 77.5% no-memory baselines while a compute-matched Retry baseline stays flat.

arXiv:2604.18206Read explainer
Memory Architecture

ActiveMem: Dynamic Latent Memory Trees for Long-Horizon Agents

Song-Li Wu, Jingyi Wang et al.

arXiv 2026 · 2026

ActiveMem organizes experiences with a Hierarchical Latent Memory Tree, Dual-Head Memory Controller, Latent Injection Head, and Tree Action Head to build dependency-aware memory paths. On ALFWorld with Qwen3-8B, ActiveMemGRPO reaches 95.57% vs MemGenGRPO’s 90.60%, while also boosting TriviaQA from 80.65% to 87.46%.

arXiv:2609.33244Read explainer

Questions about this paper?

Paper: PersMem: Internalizing Personality into Dual-Pathway Memory for LLM Agents

Answers use this explainer on Memory Papers.

Checking…