MemT-1.0, an open-source native-memory large model, cuts 256K memory usage to a constant 40M

Holy crap guys, this is seriously awesome—it feels like the most frustrating seven-year amnesia era since large models were born might finally be coming to an end!!

I'd even go so far as to say something controversial: this is probably the sexiest, yet most industry-critical, underlying breakthrough in large-model architecture evolution this year.

The just-open-sourced MemT-1.0 has, for the first time, welded a genuine biological-grade hippocampus into the very bottom of the Transformer.

The Paleolithic era of brute-forcing GPUs and propping up long context with several GB of KV Cache has officially been smashed through today by this open-source native-memory large model.

The 66-page technical report and open-source weights of up to 288B they released directly throw down a set of engineering bills that hit like a dimensional strike:

After turning memory into a native state within the forward computation flow, the VRAM overhead under a 256K ultra-long history plunges from 5.06 GB to a constant 40 MB, cutting the KV cache by more than a full 120x, and the first-token latency stays steadily at just over 100 milliseconds no matter how long the history gets.

What first breaks through the traditional path is its complete reconstruction of the model's forward computation flow.

Previously, to make AI remember context, everyone could only grit their teeth and pile on long text or bolt on external vector databases, at the cost of blowing up VRAM and losing multi-step reasoning causality. It runs a native memory module in parallel alongside the standard network; the moment input arrives, the erase gate, write gate, and forget gate operate automatically like a biological hippocampus—no longer retaining all the dead text, but digesting information in real time into a dynamic internal memory matrix.

Even more striking is the dimensional strike it delivers in real agent scenarios. The team wrote over a thousand code lessons accumulated from historical error correction directly into the internal native memory state, occupying zero system-prompt Token budget.

As a result, on standard code test sets, the hardcore metric of passing three consecutive times directly improved by 8%, with stability even surpassing the approach of stuffing in plaintext prompts.

This means experience is no longer a rigid set of rules the model must re-read and re-understand each time, but has been directly internalized into its instinctive intuition—truly awesome!

I think this is the most fundamental physical logic on the path to true recursive self-improvement: parameters only define a model's factory lower bound, but only persistently internalized native memory can open up room for compounding growth in intelligence.

Previously, the models we bought were one-off computation tools that burned out after a chat; from today on, models are becoming self-evolving assets that understand the business better the more they're used and become more valuable the more they're used.

When building agents or long-conversation applications, what gives everyone the biggest headache: the expensive VRAM and stuttering latency brought by long context, or the model's amnesia that forgets business rules the moment it turns around? For me, just compressing 256K VRAM by 120x with constant latency alone is enough to instantly revive countless long-horizon businesses that previously didn't pencil out hh~