memt-1.0 architecture breakdown gated deltanet-2, mixture-of-storage, and 66x smaller storage footprint for agent memory [D]

been digesting the details behind memt-1.0's memory-native architecture recently, and the engineering tradeoffs between full context kv caching and internalized parametric memory are pretty wild.

instead of using append-only prompt stacking or external rag pipelines, memt-1.0 bakes memory directly into the model computation layer using gated deltanet-2 with a stable token-level memory update mechanism within each turn, combined with a mixture-of-storage (mos) layer to scale capacity.

internalizing procedural experience: standard agents retrieve past execution logs as raw text, which bloats the context window. memt-1.0 continuously internalizes retrieved procedural experience into model weights/states during multi-turn agentic tasks.

efficiency at 128k context: at a history length of 128k tokens, memt-1.0-nano delivers roughly 1.86x end-to-end inference speedup and a 66x reduction in persistent storage compared to a full-context kv cache setup.

general capability preservation: training parametric memory layers usually degrades base llm capabilities, but their two-stage training pipeline (continued pre-training + post-training across 34b tokens) keeps general reasoning competitive while outperforming existing parametric memory baselines.

for those building long-horizon agents, is internalizing state via architecture-level update layers (like gated deltanet / mos) the endgame to escape kv cache memory bottlenecks?