EverMemOS unveils 4B model, 100M-token memory

This is the right direction for agent memory: not bigger prompts, but a memory layer that behaves like infrastructure.

• –MSA internalizes retrieval into the model, which could reduce the usual RAG mismatch between search and generation.
• –The 100M-token claim depends on KV-cache compression plus CPU/GPU offloading, so latency and hardware cost will matter as much as benchmark scores.
• –EverMind is productizing the research with a self-hosted repo, API docs, demos, evaluation scripts, and cloud positioning, which makes it easier to test than most long-context papers.
• –If the benchmark gains survive messy, changing data, EverMemOS could become a serious alternative to Mem0, Zep, and other agent-memory stacks.