Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 67 of the 69 references it cites were found in a registry or an index; 10 of the 12 sources its claims cite were checked for support.

Enhancing memory retrieval in generative agents through LLM-trained cross attention networks

Open access
Hong, Chuanyang, He, Qingyun · 2025 · Frontiers in Psychology

Abstract

Introduction The surge in the capabilities of large language models (LLMs) has propelled the development of Artificial General Intelligence (AGI), highlighting generative agents as pivotal components for emulating complex AI behaviors. Given the high costs associated with individually training LLMs for each AI agent, there is a critical need for advanced memory retrieval mechanisms to maintain the unique characteristics and memories of individual AI agents. Methods In this research, we developed a text-based simulation of a generative agent world, constructing a community with multiple agents…

Overview

CoChat Summary

## TL;DR

The paper proposes ACAN, a cross-attention-based memory retrieval network for generative agents, trained with LLM feedback against a weighted memory retrieval baseline. In simulations with eight agents, ACAN improved memory quality and event attendance compared with the baseline.

## Key findings

  • In a 16-hour/day simulation, the authors generated 1,280 unique training entries from 8 agents over 3 consecutive days; test simulations produced 435 data entries from 8 agents over 1 day.
  • In the memory-quality evaluation, ACAN scored 5.94 ± 1.66 versus 5.05 ± 1.88 for the WMR baseline; the difference was significant (T = 7.44, P = 2.42 × 10−13).
  • In the invitation/attendance experiment over 10 days with 100 invitation and attendance records, ACAN agents attended 32.6% of invitations on average versus 24.6% for WMR; the paired t-test gave T = 11.31, P = 0.00035.
  • The model’s training loss reportedly fell from 1.5 to 0.12.

## Methods

  • The authors built a text-based generative agent simulation world with 8 predefined agents (Farmer, Grocer, Doctor, Mayor, Chef, Hunter, Blacksmith, Carpenter) and multiple locations; interactions were simulated using ChatGPT (GPT-3.5-turbo).
  • For each decision step, they stored the agent’s current state, memory bank, and retrieved memories, embedded with text-embedding-ada-002; the training set included current state q, memory bank M_q, and WMR-retrieved memories m_r′ ranked by Recency, Importance, and Relevance.
  • They trained an Auxiliary Cross Attention Network (ACAN) with Adam (lr = 0.001, batch size = 16), using LLM-based scoring of ACAN-retrieved memories versus baseline-retrieved memories to define the loss; retrieval output was the top five memories ranked by attention weights.

## Limitations

  • The paper states that reliance on LLMs for training and evaluation increases computational demands and cost, and that LLM feedback can slow training and limit scalability.
  • The paper states that the LLM-based evaluation may not generalize across different configurations or domains, and that using LLMs instead of human assessment may affect rigor and objectivity.
  • Evident from the text: the empirical validation is based on a small simulated setup with 8 agents in a text-only environment, so broader real-world generalization is not established.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3389/fpsyg.2025.1591618

Cites

67 of 69 references verified2 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Illustration of the simulated world and memory retrieval process.
Figure 1 Illustration of the simulated world and memory retrieval process.CC BY · PubMed Central
Figure 2 Discrete time cycle of agent activities in the simulated community. The diagram illustrates the daily schedule from 6:00 AM to 9:00 PM, delineating the key activities: planning, action, and movement, underpinned by continuous perception.
Figure 2 Discrete time cycle of agent activities in the simulated community. The diagram illustrates the daily schedule from 6:00 AM to 9:00 PM, delineating the key activities: planning, action, and movement, underpinned by continuous perception.CC BY · PubMed Central
Figure 3 Training loss curve over epochs.
Figure 3 Training loss curve over epochs.CC BY · PubMed Central
Figure 4 Score distribution of memory retrieved by ACAN and weighted memory retrieval (WMR).
Figure 4 Score distribution of memory retrieved by ACAN and weighted memory retrieval (WMR).CC BY · PubMed Central

Verification

Claims the paper makes

7 Valid · 2 Not valid · 2 Not checked · 39 Own results and interpretations

13 more claims

Create a free account to see all 50 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 18
  • It may state two separate claims as one. 18

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 61 references

Research works printed but not linked: 2

  1. Musacchio, 2024, Adapting large language models to narrative content, Complexity, 6, 8
  2. Zalta, 1995, Stanford Encyclopedia of Philosophy

Cited by

See all 12

Metadata

Authors
Hong, C., & He, Q.
Affiliations
Southwestern University of Finance and Economics (CN)
Venue
Frontiers in Psychology
Published
2025-05-07
Volume
16
Pages
1591618-1591618
ISSN
1664-1078 (ISSN-L)
OpenAlex
W4410182127
Cited by
7
References
61
Citations by year
2026: 4 · 2025: 3
FWCI
2.80
Type
journal article
Language
en
Open access
Locations
  • Frontiers in Psychology (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Psychiatry, Mental Health, Neuroscience
  • Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
Keywords
generative AI agents, memory retrieval, cross-attention network, large language models, artificial general intelligence, loss function design, agent-based modeling, behavioral consistency

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Enhancing memory retrieval in generative agents through LLM-trained cross attention networks. https://app.cochat.ai/papers/10.3389/fpsyg.2025.1591618
The paper
Hong, C., & He, Q. (2025). Enhancing memory retrieval in generative agents through LLM-trained cross attention networks. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2025.1591618

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.