Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 13·5 min read
Consolidator: Learning Persistent Routed Memory Across Context Boundaries
TL;DR
Consolidator enhances memory retention in Phasor Memory Networks by transforming short-term memory (STM) into long-term memory (LTM) without replaying source tokens. This method significantly improves recall of updated mappings, achieving a recall increase from 44.38% to 87.02%.
✦ Why It Matters
Implementing Consolidator can enhance memory recall in your AI models today.
Key Takeaways
- Explore using Consolidator for memory tasks in AI models.
- Consider the impact of memory access states on model performance.
- Evaluate the effectiveness of LTM in your applications.
- Train only a small number of parameters for significant improvements.
How It Works
Consolidator processes short-term memory (STM) by transforming it into long-term memory (LTM) without replaying the original tokens. After each consolidation, the model clears the KV cache and STM, allowing the retained LTM to influence future memory access through a hierarchical router.
This mechanism ensures that the updated mappings are effectively recalled in subsequent queries.