Multimodal Agent MemorySubmit a paper

MemShot

Memory Shot for Long-Term Dialogue

Chunyi Peng, Haidong Xin, Xuanshuo Sheng, Xin Dai, Zhenghao Liu, Shuo Wang, et al. · 2026-05-30

Read the paperPDF

Management
T1 Static Memory
Memory is built once and reused without change during deployment.
Primary modality
Image
Also: text
Representation
Source- and modality-specific

How experience becomes memory

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

  1. Memory shots: rendered images of contiguous dialogue spans with a session/timestamp header and a speaker-aware chat region

    1. Represented asRetained perceptual source evidenceSource- and modality-specific
    2. Stored inVector-indexed collection
    3. Retrieved bySimilarity search
    4. Source evidenceSource directly available
    Dialogue spans are rendered into images rather than compressed into text: a header region carries 'session-level metadata such as a session identifier and timestamp' and a chat region lays out 'dialogue as a speaker-aware chat stream' with consistent per-speaker positions and styles. Turn-pairs are packed until the 948x768-pixel canvas is full, and each new unit prepends 'the last two turn-pairs from the previous unit' to preserve continuity. The images are the memory items: they are embedded with Qwen3-VL-Embedding-8B, the top-10 units are retrieved by embedding similarity, and the retrieved images themselves are fed to the model, which uses 'its internal visual reasoning capabilities to associate key episodes'. The dialogue text is preserved verbatim inside the image, so the source is directly available downstream.Full paper (HTML v1): 3.1 (memory construction, Eq. 3); 3.3 Visual memory unit construction (Eq. 9); 4 Experimental setup; A.3 Implementation details · HTML full text · Table 5 (rendering parameters); Figure 4 (retriever robustness)

Related systems