Pensieve
Memory-QA: Answering Recall Questions Based on Multimodal Memories
Hongda Jiang, Xinyuan Zhang, Siddhant Garg, Rishab Arora, Shiun-Zu Kuo, Jiayang Xu, et al. · 2025-09-22
- Management
- T1 Static Memory
- Memory is built once and reused without change during deployment.
- Representation
- Hybrid
- Text and structured metadata remain linked to the original image entry; the VLM configuration can directly consume retrieved images.
How experience becomes memory
Each path is one memory component, traced from what is stored to how it is found again.
multimodal memory entries
- Represented asRetained perceptual source evidenceSource- and modality-specific
- Stored inFlat sequential records
- Retrieved bySimilarity search, Metadata / spatiotemporal retrieval
- Source evidenceSource directly available
Each stored memory entry contains an image snapshot plus its invocation, timestamp, and location metadata.Full paper: Memory Retrieval · PDF p. 4
offline memory augmentation
- Represented asPerception-to-text records & summariesCross-represented explicit
- Stored inVector-indexed collection
- Retrieved bySimilarity search, Metadata / spatiotemporal retrieval
- Source evidenceSource linked & recoverable
OCR, QA-guided visual descriptions, and completed invocation text augment each source memory before retrieval.Full paper: Memory Retrieval · PDF p. 4 · Figure 2
memory metadata
- Represented asSpatial, temporal & cross-modal structured memoryCross-represented explicit
- Stored inGraph / relational store
- Retrieved byMetadata / spatiotemporal retrieval
- Source evidenceSource linked & recoverable
Date match, recency, location similarity, and multimodal embedding similarity are fused to rank memory entries.Full-paper method / memory representation and retrieval · PDF p. 5 · Figure 3
Related systems
- FinSight (Code Agent with Variable Memory)T2 · Image
- GeoForgeT3 · Image
- VideoAgentT1 · Video & streaming
- AtlasVAT1 · Image
- GRGA (Graph-based Retrieval-Generation Agent)T2 · Audio & speech
- HiKEYT1 · Image