Memory Decoder at Scale: 6.9B Parameter Parametric Long-Term Memory
July 29, 2026
Memory Decoder scales parametric long-term memory to 6.9B parameters trained on 300B tokens. The architecture uses a distributed Faiss pipeline and sparse, batch-wise kNN loading to overcome the indexing and retrieval bottlenecks encountered at scale.
HOW THIS AFFECTS YOU
●
builderAllocating parameters specifically to memory may provide a better performance tradeoff than simply scaling base model size.
●
researcherThe distributed Faiss pipeline offers a solution for scaling kNN retrieval in large-scale memory models.