Long-Context Memory & Inference Breakthroughs
Key Questions
What is SAM in long-context agent memory?
SAM is a state-adaptive memory system designed for long-horizon reasoning agents. It demonstrates strong results on benchmarks such as BrowseComp and HLE.
How does EverMemOS improve AI memory?
EverMemOS introduces a biologically-inspired three-phase memory lifecycle using MemCells, MemScenes, and Reconstructive Recollection. It significantly outperforms flat memory systems on long-context benchmarks.
What is δ-mem and its compression benefit?
δ-mem compresses context representations into an 8x8 matrix for efficient long-context handling. It contributes to broader advances in memory and inference efficiency.
GoLongRL; inference scaling 8B-671B; DiGraphHal-Bench; OScaR KV cache; MinT serving; Multi-Stream LLMs; Gated DeltaNet-2. New: δ-mem compresses context to 8x8 matrix; DeepSeek-V4 9.5x KV cache reduction. Latest: SAM state-adaptive memory for long-horizon reasoning agents (strong on BrowseComp/HLE). Also FluxMem graph-based agent memory (SOTA on LoCoMo/Mind2Web/GAIA). New: EverMemOS biologically-inspired three-phase memory lifecycle (MemCells, MemScenes, Reconstructive Recollection) significantly outperforms flat memory systems on long-context benchmarks.