Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MMA: Multimodal Memory Agent

Yihao Lu, Wanru Cheng, Zeyu Zhang, Hao Tang

TL;DR

The paper tackles the reliability of memory-augmented long-horizon agents by addressing retrieval traps and overconfident errors. It introduces MMA, a memory-agent with a Confidence Module that assigns each retrieved memory a reliability score based on source credibility, temporal decay, and cross-memory consensus, enabling reweighting and abstention when evidence is insufficient. It also presents MMA-Bench, a benchmark designed to stress test belief dynamics with controlled reliability priors and cross-modal contradictions, revealing phenomena like the Visual Placebo Effect in RAG-based systems. Across FEVER, LoCoMo, and MMA-Bench, MMA achieves comparable accuracy with substantially improved stability and calibrated abstention, demonstrating the practical value of epistemic prudence in memory-augmented decision making.

Abstract

Long-horizon multimodal agents depend on external memory; however, similarity-based retrieval often surfaces stale, low-credibility, or conflicting items, which can trigger overconfident errors. We propose Multimodal Memory Agent (MMA), which assigns each retrieved memory item a dynamic reliability score by combining source credibility, temporal decay, and conflict-aware network consensus, and uses this signal to reweight evidence and abstain when support is insufficient. We also introduce MMA-Bench, a programmatically generated benchmark for belief dynamics with controlled speaker reliability and structured text-vision contradictions. Using this framework, we uncover the "Visual Placebo Effect", revealing how RAG-based agents inherit latent visual biases from foundation models. On FEVER, MMA matches baseline accuracy while reducing variance by 35.2% and improving selective utility; on LoCoMo, a safety-oriented configuration improves actionable accuracy and reduces wrong answers; on MMA-Bench, MMA reaches 41.18% Type-B accuracy in Vision mode, while the baseline collapses to 0.0% under the same protocol. Code: https://github.com/AIGeeksGroup/MMA.

MMA: Multimodal Memory Agent

TL;DR

The paper tackles the reliability of memory-augmented long-horizon agents by addressing retrieval traps and overconfident errors. It introduces MMA, a memory-agent with a Confidence Module that assigns each retrieved memory a reliability score based on source credibility, temporal decay, and cross-memory consensus, enabling reweighting and abstention when evidence is insufficient. It also presents MMA-Bench, a benchmark designed to stress test belief dynamics with controlled reliability priors and cross-modal contradictions, revealing phenomena like the Visual Placebo Effect in RAG-based systems. Across FEVER, LoCoMo, and MMA-Bench, MMA achieves comparable accuracy with substantially improved stability and calibrated abstention, demonstrating the practical value of epistemic prudence in memory-augmented decision making.

Abstract

Long-horizon multimodal agents depend on external memory; however, similarity-based retrieval often surfaces stale, low-credibility, or conflicting items, which can trigger overconfident errors. We propose Multimodal Memory Agent (MMA), which assigns each retrieved memory item a dynamic reliability score by combining source credibility, temporal decay, and conflict-aware network consensus, and uses this signal to reweight evidence and abstain when support is insufficient. We also introduce MMA-Bench, a programmatically generated benchmark for belief dynamics with controlled speaker reliability and structured text-vision contradictions. Using this framework, we uncover the "Visual Placebo Effect", revealing how RAG-based agents inherit latent visual biases from foundation models. On FEVER, MMA matches baseline accuracy while reducing variance by 35.2% and improving selective utility; on LoCoMo, a safety-oriented configuration improves actionable accuracy and reduces wrong answers; on MMA-Bench, MMA reaches 41.18% Type-B accuracy in Vision mode, while the baseline collapses to 0.0% under the same protocol. Code: https://github.com/AIGeeksGroup/MMA.
Paper Structure (38 sections, 8 equations, 11 figures, 9 tables)

This paper contains 38 sections, 8 equations, 11 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. The Proposed Method And Benchmark
  4. Overview
  5. Multimodal Memory Agent (MMA)
  6. MMA-Bench
  7. Experiments
  8. Robustness on Standard Benchmarks
  9. Results on MMA-Bench
  10. Evolutionary Cognitive Analysis
  11. Ablation Study
  12. Conclusion
  13. Extended Related Work
  14. Memory Architectures And Control Policies.
  15. Compressed And Synthesized Memory Representations.
  16. ...and 23 more sections

Figures (11)

  • Figure 1: Retrieval Trap Case Study. MIRIX fails by retrieving the high-similarity but irrelevant Memory B. MMA correctly identifies the credible Memory A using multi-dimensional reliability signals.
  • Figure 2: MMA Framework. The Confidence Module reweights retrieval via source reliability, temporal decay, and network consensus to modulate reasoning and abstention.
  • Figure 3: MMA-Bench evaluation framework. The benchmark integrates cross-modal consistency analysis, risk-aware betting, and a $2\times2$ logic matrix for trust conflicts. Performance is assessed through fundamental QA and a 3-step belief probe.
  • Figure 4: Detailed Dynamics Analysis. (a) Step-wise belief revision; (b) Risk-adjusted scores highlighting visual noise sensitivity; (c) Gap analysis between retrieval accuracy and calibration.
  • Figure 5: Evolutionary Logic Spectrum. Tracing performance from Foundation Model to MMA. MMA restores agency (Activation) and buffers inherited visual bias (Placebo Effect).
  • ...and 6 more figures