
2
modele
1
języki
EverOS is a long-term memory operating system for AI agents, developed by EverMind AI (San Mateo, CA). It addresses the architectural limitation of stateless large language models by providing a structured memory infrastructure that persists, consolidates, and retrieves knowledge across sessions. The platform is available both as an open-source self-hosted deployment (GitHub) and as EverOS Cloud at everos.evermind.ai.
EverOS operates through a four-layer architecture: the Agentic Layer (task understanding, planning, execution), the Memory Layer (long-term storage and retrieval), the Index Layer (embeddings, key-value pairs, knowledge graph indexing), and the API/MCP Interface Layer (integration with external enterprise systems). Memory management follows an engram-inspired three-phase lifecycle: Episodic Trace Formation (converting dialogue streams into structured MemCells), Semantic Consolidation (organizing MemCells into thematic MemScenes), and Reconstructive Recollection (MemScene-guided agentic retrieval composing necessary context for downstream reasoning).
EverOS features four key innovations: (1) Self-Evolving Agent Memory — a Skills Evolution Engine that automatically distills reusable skills (SOPs) from completed tasks via agent case extraction, semantic clustering, and skill emergence; (2) mRAG Hybrid Retrieval Architecture — native multimodal memory ingestion (PDFs, images, Word docs, spreadsheets, URLs) through a single API, fusing dense semantic vectors, sparse keyword matching, and multimodal alignment; (3) HyperMem Architecture — a hypergraph memory network (accepted at ACL 2026) replacing flat vector databases to capture multi-hop, cross-temporal entity relationships with ultra-low latency; (4) RESTful API with MCP interface and EverOS Cloud Playgrounds (Coding Playground integrated with Google Colab, Chat Playground for visual memory comparison).
EverCore (the underlying engine) achieves 93.05% accuracy on LoCoMo, 83.00% on LongMemEval, 90.04% on HaluMem, and SOTA performance on PersonaMem v2. The Skills Evolution Engine yielded a 234.8% relative increase in task success rate for software engineering problems (27B model, EvoAgentBench). The underlying research is published as arXiv:2601.02163 (EverMemOS).
Supported storage engines include SQLite, PostgreSQL, and vector databases with embeddings (FAISS, Milvus, pgvector). The platform supports multi-tenant memory IDs and is compatible with LangGraph, Haystack, and other agent frameworks. It integrates with any LLM API (OpenAI, Qwen, Llama, local models via API wrapper) and embedding services. All memories are stored as transparent JSON objects with timestamped, persistent episodes.
EverOS is designed for personalized AI assistants, customer service agents requiring continuous contextual understanding, multi-user collaboration and knowledge retention, research and analysis (building knowledge bases from conversations), and educational tools adapting to learning patterns.
Modele cenowe
Limity zasobów
SLA i wsparcie