Two-part memory improves recall in multimodal programs with far fewer tokens
A memory system that expands recall across past episodes and shows only helpful details from each memory.
- Publication
- arXiv
- Stage
- Preprint
- What we read
- Summary of the paper
- Authors
- Haozhen Zhang, Bohan Liu, Jianzhu Bao, Qingyuan Liu, Haodong Yue, Xiao Li, Zi Pong Lim, Wenya Wang
- Universities and research institutions
- Nanyang Technological University, Tsinghua University, Sun Yat-sen University, Aumovio Singapore Pte. Ltd
Institution metadata: Original paper ↗
What they did and found
Researchers tested a two-part memory approach for programs that handle text and pictures. It first expands recall across past episodes, then shows only the most useful details from each memory. Results: better task performance with far less token use.
Why it matters
This approach can cut data needs while keeping accuracy in long text–image chats, helping faster and cheaper responses in real-world use.
The memory groups chats and visuals into small chunks and links past episodes to find related memories that help answer current questions.
For each memory, the system selects only the most helpful text and image details to keep answers clear and costs down.
What remains uncertain
Study uses a preprint with three benchmarks and two base models; results may not generalize to all tasks or systems.
Original sources · 1
- HiVe-Mem: Interleaving Horizontal Memory Expansion and Vertical Evidence Composition for Multimodal Agents ↗arXiv · 2026-11-09
Check the original paper for its authors, methods, version and access terms.
What is your take?
Ask a question, add useful context or share a different perspective. Keep the conversation respectful and grounded.