Δ-Mem: Efficient Online Memory for Large Language Models
4 hours ago
- #large-language-models
- #artificial-intelligence
- #memory-mechanism
- Proposed δ-mem, a lightweight online memory mechanism for large language models.
- Uses a fixed-size state matrix updated by delta-rule learning to compress past information.
- Generates low-rank corrections to attention computation during generation.
- Achieves performance improvements, e.g., 1.31× on MemoryAgentBench, with minimal memory overhead.
- Works without full fine-tuning, backbone replacement, or explicit context extension.