Multimodal Agent MemorySubmit a paper

M2A

M2A: Multimodal Memory Agent with Dual-Layer Hybrid Memory for Long-Term Personalized Interactions

Junyu Feng, Binxiao Xu, Jiayi Chen, Mengyu Dai, Cenyang Wu, Haodong Li, et al. · 2026-02-07

Read the paperPDF

Management
T2 Dynamic Memory
Memory changes during deployment through a stable management policy.
Primary modality
Image
Also: text, conversation
Representation
Hybrid
Editable semantic observations and their image fields remain linked by evidence_ids to immutable raw messages.

How experience becomes memory

Each path is one memory component, traced from what is stored to how it is found again.

  1. immutable raw conversation and image log

    1. Represented asRetained perceptual source evidenceSource- and modality-specific
    2. Stored inHierarchical multi-tier store, Flat sequential records
    3. Retrieved byPlanner / tool-assisted retrieval
    4. Source evidenceSource directly available
    The append-only RawMessageStore preserves complete multimodal messages in chronological order and remains addressable through evidence identifiers.Full paper: 4.1. Framework Overview · PDF p. 4 · Figure 2

  2. editable high-level semantic observations

    1. Represented asPerception-to-text records & summariesCross-represented explicit
    2. Stored inHierarchical multi-tier store, Vector-indexed collection
    3. Retrieved bySimilarity search, Planner / tool-assisted retrieval
    4. Source evidenceSource linked & recoverable
    Semantic entries combine high-level descriptions, image captions, associated images, and pointers to supporting raw-message ranges; dense, sparse, and visual rankings are fused.Full paper: 4.1. Framework Overview · PDF p. 4

How memory changes (T2 operations)

  • Update · Prompted LLM/VLM record update

    MemoryManager analyzes prior entries and executes create, delete, or modify operations after ChatAgent requests an update. For M2A, this specifically supports LLM-mediated memory editing of immutable raw conversation and image log.

  • Revision · LLM-based conflict resolution

    MemoryManager compares a new interaction with related memories and removes or replaces outdated and conflicting semantic entries. For M2A, this specifically supports LLM-based resolution of conflicting records of immutable raw conversation and image log.

Related systems