EverMemOS

EverOS — also branded EverMemOS — is a memory operating system for AI agents, built by EverMind on the EverCore engine. It targets a specific failure: LLMs are stateless, so every session restarts from zero and anything past the context window is gone. EverOS gives agents persistent, proactive, cross-session recall that survives days, platforms, and team handoffs. The architecture has four moving parts worth knowing. Self-evolving Skill Memory records agent trajectories as Cases, distills repeated patterns into reusable Skills, and lets agents improve from prior experience. The mRAG layer handles multimodal ingestion and cross-modal retrieval, parsing PDFs, images, Word documents, spreadsheets, presentations, emails, HTML pages, text files, and URLs through one API. Memory Bank is the transparency layer: user, group, and agent memory are separated, inspectable, and editable — including the generated Skills themselves. Temporal tracking distinguishes a current fact from an outdated one when pricing, roles, or preferences change. Recent engineering work sharpened retrieval rather than breadth. Multi-round retrieval adds a small decider model that stops the search early — 1.95 episodes and 23% of the k=20 prompt reach 97.4% of peak accuracy, taking first place on four memory benchmarks, while a small decider model stops the search. CRAFT compresses video tokens by merging globally similar tokens and gating position-aware weights, retaining 96.8% of backbone accuracy at roughly 8x compression. SkillCorpus filters ~821,000 crawled skills down to 96,401 audited ones paired with a fine-tuned retrieval stack, and HarnessBank pairs LLM-driven diagnosis with deterministic credit assignment for harness changes. This is infrastructure for engineering teams, not a consumer app. Publlished benchmarks — LoCoMo 93.05%, LongMemEval 83.00%, HaluMem 93.04% — are open to reproduction. Against lighter memory layers like Mem0, EverOS trades setup effort for multi-agent coordination,