Memorilla Shrinks LLM Memory to 16 Tokens

Memorilla makes a compelling case that learned, query-conditioned compression can outperform simply stuffing more context into an LLM.

• –Precomputed document embeddings keep storage separate from the frozen language model, enabling scalable external memory.
• –It outperforms Mem0, MemoRAG, and other baselines on five of eight benchmarks.
• –The fixed token budget could make long-term agent memory cheaper and more predictable.
• –Results deserve scrutiny because the memory module receives task-specific training while several baselines are evaluated off the shelf.
• –The open-source release includes training, evaluation, datasets, and TextWorld reinforcement-learning support.