memt-1.0's memory-native architecture: 1.86x faster, 66x smaller storage at 128k context
Mrs-Grumpy23 · reddit · 2026-10-25
A detailed breakdown of memt-1.0's "memory-native" architecture: instead of append-only prompt stacking or external RAG, it internalizes memory into the computation layer via gated DeltaNet-2 with stable token-level memory updates, plus a mixture-of-storage (MoS) layer for capacity.
- Internalizing procedural experience: agents normally retrieve past execution logs as raw text, bloating context; memt-1.0 continuously bakes retrieved experience into weights/states during multi-turn tasks.
- Efficiency: at 128k-token history, memt-1.0-nano achieves 1.86x end-to-end inference speedup and 66x less persistent storage vs a full-context KV cache setup.
- Capability preservation: a two-stage pipeline (continued pre-training + post-training over 34B tokens) keeps general reasoning competitive while beating existing parametric-memory baselines.
- The attached chart shows comparisons against MemOS, DenseRAG, Graphiti, MemoryLLM and more across six benchmarks (LoCoMo, MemoryAgentBench, ImplicitMemBench, MemBench, PerMemSafe, PersonaMem).
The author asks whether architecture-level memory internalization is the endgame for long-horizon agents escaping KV cache bottlenecks.
Related event: MemT-1.0 Bakes RAG-Level Memory Into Model Parameters(2 posts)→
More from coding & agent
- How this open-source contributor consistently finds impactful performance bugs: dissect subsystems and experiment — Abhishekcur · 2026-10-25
- Dev trains a 4B coding model that can also generate 3D scenes — ConfidentDinner6648 · 2026-10-25
- Stop checking what your agent said — check what it actually changed — Key-Register-1969 · 2026-10-25
- Dev builds a cozy Japanese-learning village game entirely with Claude Opus — thetinytrex · 2026-10-25
- Base44 adds Entri integration: connect custom domains to your apps in a couple of clicks — tristanbob · 2026-10-25
- Grok Bot burns a week of Cursor Ultra quota in one day for a $200/mo user — alexcovo_eth · 2026-10-25