Frontier AI, out-remembered.
The benchmark is MRCR v2 8-needle: a long conversation full of requests that look alike, where the model has to resolve a reference back to the right earlier one. Native Context Memory is model-agnostic.
1M-token context
5M-token context
- 1DeepSeek-V4-Flash + Native Context Memory98.5%Adaptiq
5M is just the start.
A 5M-token context in each agent session is the foundation we build on. We believe unlimited context is achievable, and that is what we are working towards.
What Native Context Memory adds.
We ran MiMo V2.6 Flash through all ten MemoryAgentBench tasks, with and without NCM. MemoryAgentBench measures whether an LLM can remember over time: not whether it can read a long text once, but whether it can store, update and retrieve what it has been told in pieces along the way.
It keeps track of what changed.
When a fact is replaced by a newer one, MiMo on its own often misses the change. With NCM it goes with the newest.
It connects what was said at different times.
How many, how long ago, which came first: when the answer is spread over several conversations, a model alone loses the thread. NCM finds each piece and puts them together.
It stays exact, however much there is to remember.
On its own, MiMo’s answers vary from one long text to the next. With NCM they hold steady, because the answer comes from the lines themselves.
MemoryAgentBench
Conflict resolution
Accurate retrieval
Test-time learning
Long-range understanding
Tokens used
MiMo alone used 3.5 times more new input tokens than MiMo with NCM: 179.0 million against 51.5 million across all 2,071 questions.
| MiMo alone | MiMo + NCM | |
|---|---|---|
| New input tokens | 179.0M | 51.5M |
| Input read from cache | 639.7M | 600.0M |
| Total input | 818.7M | 651.5M |
| Output tokens | 0.60M | 5.25M |
| New input tokens per question | about 86,400 | about 24,900 |
| Total cost (OpenRouter) | $27.01 | $10.36 |