When AI Starts Rewriting Itself: EverMind’s Raven Agent Redefines Digital Life

Stateless limitation of current AI assistants

Most AI assistants operate as stateless machines: when a conversation window is closed, all context is discarded, preventing the system from retaining experiences or growing over time.

EverOS memory operating system

EverOS provides a four‑layer bio‑inspired architecture—Agent layer, Memory layer, Index layer, and Interface layer—designed to embed long‑term, user‑centric memory into an agent’s cognition.

Between May and June 2026 EverOS versions 1.0.0, 1.0.1 and 1.1.0 were released, quickly surpassing 10 k GitHub stars (mem0 accumulated 7 k stars in the same period).

EverOS achieves token consumption of roughly 1/10 of traditional solutions while preserving full‑context accuracy.

Version 1.1.0 adds explicit categories—User Memory, Agent Memory, and Knowledge Wiki—and introduces a “Reflection” mechanism that periodically consolidates insights.

Key research contributions

MSA (Memory Sparse Attention) : an end‑to‑end trainable sparse‑attention mechanism extending context length to 100 million tokens with less than 9 % performance degradation; topped HuggingFace Daily Papers on release day.

HyperMem : a hypergraph‑based hierarchical memory architecture achieving 92.73 % SOTA on the LoCoMo benchmark; received an ACL 2026 Oral Presentation honor.

EverMemOS : a self‑organizing memory OS for structured long‑term reasoning; selected for the ACL 2026 main conference.

Multi‑party Collaborative Long‑Term Memory Evaluation : fills an industry gap; awarded a KDD 2026 Oral Presentation.

Raven Agent

Raven is a self‑improving agent harness built on EverOS. Its core values are Proactive, Improving, and Personalized.

Memory is bidirectional: Raven continuously updates a deep user model while extracting self‑improvement signals from each interaction.

Raven ships with 100,000 rigorously evaluated skills covering everyday productivity to specialized vertical domains; the skill set is dynamically assessed, retired, or combined based on real‑world usage.

During idle periods Raven can rewrite its own logic and fine‑tune model weights via the EverBrain user‑side memory model, laying groundwork for the L4 “Autonomous Digital Life” stage.

Digital‑life stages

L1 – Role‑based functional agents: stateless, limited to the current dialogue window.

L2 – Memory‑augmented interactive agents: basic long‑term memory, multi‑step planning, API integration.

L3 – Self‑Improving Cognitive Agents: reinforcement learning, reflection, dynamic weight updates (embodied by Raven).

L4 – Autonomous Digital Life: agents initiate tasks, manage their own memory, and act without explicit user prompts.

Pluggable architecture and developer ecosystem

Memory, proactive engine, and tool routing are independent modules; developers can replace any component without modifying the core framework.

Future “Raven Builder” will enable developers to define custom agents for arbitrary scenarios and share them instantly.

Full‑stack ecosystem

Bottom: EverOS (open‑source Memory OS).

Model: EverBrain (user‑side memory model for personalized inference).

Agent: Raven (self‑improving agent harness).

User: EverMe (personal digital‑twin platform).

Installation

curl -fsSL https://raven.evermind.ai/install.sh | bash

GitHub repository: https://github.com/EverMind-AI/Raven