Described in arXiv:2610.21375 from Stanford, UPenn, MIT and Vector Institute collaborators, Memorilla uses a frozen document encoder to build a persistent embedding matrix, conditions learned memory slots on the query embedding, attends over the full matrix with Perceiver-style blocks, and projects slots into the frozen decoder's embedding width so the LLM consumes memory like ordinary tokens. The default supervised configuration in the paper fixes K=16 memory tokens for a frozen Qwen3-8B decoder and compares against Mem0, MemoRAG, xRAG and full-context baselines on eight text benchmarks plus a TextWorld agent ablation.