MemT-1.0 claims RAG-level accuracy in parametric memory: 1.86x faster inference at 128K history
Alternative-Pay-447 · reddit · 2026-10-25
A Reddit post walks through memt-1.0, an architecture that internalizes procedural experience into model state via gated DeltaNet-2 and mixture-of-storage (MoS), aiming to escape the 'slow & bloated RAG vs. light but dumb fine-tuning' trade-off.
- At 128K history it reports 1.86x end-to-end inference speedup and 66x less persistent storage, avoiding KV cache blowups on long multi-turn sessions.
- In the parametric memory tier it scores 73.5 on MemBench and 65.1 on LocoMo, versus the 20–30% typical of legacy parametric approaches like doc-to-lora and MemoryLLM.
- A two-stage training pipeline (34B tokens) reportedly preserves base-model reasoning while embedding long-term memory.
The post ends by asking whether agent builders will stick with external vector RAG or move to architecture-level memory internalization; numbers are self-reported and unverified.
More from coding & agent
- Open-source decision model Laya hits 32K phones in one month at ~33ms per call — HowDevelop · 2026-10-25
- Agents Mirror Your Org Chart Just Like Microservices Did, Engineer Argues — arpit_bhayani · 2026-10-25
- AI-built bbtree claims fastest concurrent btree; researcher calls for ambitious academic systems — tianyin_xu · 2026-10-25
- Froxel: open-source voxel editor with built-in MCP server lets AI agents model and render — roman_va · 2026-10-25
- CCC dashboard treats 'waiting on human' as a separate state from agent busy — amirfish · 2026-10-25
- Octopus OS: a 45k-line Rust kernel where agents rewrite their own harness mid-task — _AustinCalvert_ · 2026-10-25