δ-mem gives LLMs infinite online memory

δ-mem is the most promising alternative to RAG we've seen this year, turning memory into a continuous, differentiable sidecar rather than a clunky external search.

• –Delta-rule updates allow the model to learn and refine facts during a single session by prioritizing novel, relevant information
• –Compact 8x8 matrix adds near-zero overhead while boosting agentic performance by 31% on MemoryAgentBench
• –Integrated directly into the attention mechanism to eliminate retrieval-induced noise and "lost in the middle" effects
• –Works with frozen backbones, making it a viable "plug-in" for both proprietary and open-weights models
• –Constant-time complexity solves the "infinite context" problem for long-form agents and real-time assistants