138 likesNew · 11d old
Tiny memory model slashes context storage 66× while remembering
What is MemT-1.0-Nano?
MemT-1.0-Nano adds a persistent memory block to a 35-billion-parameter language model, letting it keep a small, evolving state across prompts. The design cuts the usual key-value cache size by about two-thirds for a 128-k token history.
Who it is for
If you build agents or chatbots that need long-term context, you can experiment with a model that remembers without blowing up RAM.