MemT-1.0: open memory-native LLMs cut KV storage 66x at 128K
MemTensor's Nano (36B/3B) and Flash (288B/17B) bake memory into the weights — MIT, OpenAI-compatible, and a constant-size state for long-context agents.
Copy markdown
Memory lives in the weights, not a ballooning KV cache
MemT-1.0 is a native-memory architecture (Gated DeltaNet-2 plus a Mixture-of-Storage) that holds a constant-size persistent state instead of a cache that grows with the conversation. At a 128K history, Nano reports a 66x cut in persistent storage versus a full-context KV cache — the economics of serving 100K+ token agents change once the state stops scaling with history.
Two sizes, both cheap to run
Nano is 36B total / 3B active on a Qwen3.6-35B-A3B backbone (10M dynamic memory); Flash is 288B total / 17B active on DeepSeek-V4-Flash (336M dynamic memory). With only 3B active, Nano serves on modest hardware while carrying persistent memory the backbone never had.
The speedups that hit your token bill
The longer the context, the bigger the win: ~1.86x end-to-end inference speedup at 128K, and time-to-first-token 2.44x faster at 128K and 4.19x faster at 256K. If your agent re-reads a long history every turn, that latency comes straight off your critical path.
MIT weights that drop into your stack
Weights ship under MIT; vLLM and SGLang serve the model over an OpenAI-compatible /v1 endpoint, and a session memory API (enable_memory=True) saves and restores memory state to a file. Build this weekend: a self-hosted assistant that keeps a persistent memory file across sessions instead of re-stuffing the whole history into every prompt.
Read the fine print before you switch
The gains are the authors' own evals, and they are candid about the ceiling: across 11 memory benchmarks Flash leads parametric-memory methods (avg 50.6) but text/retrieval memory systems still edge ahead on long histories that need detailed evidence. No hosted provider runs it yet — self-hosting is the only path today.