First Release | Chen Tianqiao’s Shanda Team Launches Powerful Open-Source Memory System EverMemOS
EverMemOS Sets a New Standard in AI Long-Term Memory
Date: 2025-11-16 11:59 — Location: Hong Kong, China
EverMemOS has delivered significantly better performance than previous systems on mainstream benchmarks such as LoCoMo and LongMemEval-S, establishing the new State-of-the-Art (SOTA) in AI long-term memory.
---
Introducing EverMemOS
The EverMind team has officially launched EverMemOS, a world-class long-term memory operating system for AI agents. Designed to serve as the foundational data infrastructure for next-generation intelligent agents, EverMemOS enables AI to have a lasting, coherent, and evolvable "soul".
Key Achievements
- Outstanding benchmark results:
- EverMemOS achieved 92.3% on LoCoMo and 82% on LongMemEval-S, surpassing all previous systems.
- Industry recognition:
- Considered a breakthrough product in AI infrastructure.
Links:
- Website: http://everm.ai
- GitHub: https://github.com/EverMind-AI/EverMemOS/
---
Memory Capacity — The Watershed for Next-Gen AI
The Problem
Large Language Models (LLMs) have fixed context windows, causing them to “forget” over extended tasks. This leads to:
- Broken memory and contradictory facts
- Lack of deep personalization
- Inability to retain intermediate context data
Why It Matters
Without robust long-term memory:
- Agents cannot maintain behavioral consistency
- Self-iteration is impossible
- Personalization, consistency, and initiative — key to advanced AI — remain out of reach
Industry Momentum
Products like Claude and ChatGPT have already made long-term memory a strategic feature, marking memory as the core competitive advantage in future AI applications. Yet, most existing solutions:
- Are fragmented or scenario-limited
- Fail to unify precision, speed, usability, and adaptability
---
Inspired by the Human Brain
Background
EverMind’s team hails from Shanda Group, with roots in tech innovation and investment. Their design closely mirrors biological memory processes:
- Sensory encoding
- Hippocampal indexing
- Cortical long-term storage
- Prefrontal cortex coordination
This human-brain-inspired approach aims to give AI continuity over time, enabling adaptation and long-term growth.
---
The Temporal Structure Paradigm
At the Tianqiao Institute for Brain Science’s AIAS 2025 Symposium, Chen Tianqiao outlined five pillars of discovery-based intelligence, highlighting long-term memory as the bridge between time and intelligence.
Key Insight:
Human brains operate in continuous, dynamic temporal structures, whereas AI typically works in instantaneous, static spatial structures.
EverMemOS is designed to connect time and intelligence by granting AI temporal continuity.
---
Breakthroughs in Scenario Coverage & Performance
- Scenario Coverage:
- First memory system to fully support:
- 1-to-1 conversations
- Multi-user collaboration
- Already integrated into AI Native Tanka.
- Technical Performance:
- Uses bio-inspired “engram” heuristic retrieval to achieve:
- 92.3% (LoCoMo)
- 82% (LongMemEval-S)
- — both SOTA results.
---
EverMemOS Four-Layer Architecture
Layers & Their Analogy to Brain Regions
- Agentic Layer — Task understanding & executive control (prefrontal cortex)
- Memory Layer — Long-term structured storage (cortical network)
- Index Layer — Fast association via embeddings & knowledge graphs (hippocampus)
- API/MCP Interface Layer — Links AI with external applications (sensory systems)
---
Three Core Features
1. From “Memory Database” to “Memory Processor”
EverMemOS actively applies retrieved memories to reasoning — ensuring every AI output reflects long-term user context.
2. Layered Memory Retrieval & Dynamic Organization
Memories are stored as episodic semantic units, linked dynamically to capture implicit context beyond keyword similarity.
3. Modular & Extensible Memory Framework
Supports varied memory needs:
- High-precision structured data for professional scenarios
- Emotional context & empathy for companion AI scenarios
---
From Open Source to Cloud Services
EverMemOS is available open-source on GitHub:
https://github.com/EverMind-AI/EverMemOS/
Planned cloud service version to offer:
- Enterprise-grade support
- Data persistence
- Scalable deployment
Sign up: http://everm.com for early access.
---
Mission Statement
> “We are confronting one of the most profound challenges in AI — giving machines memory and opening the door to higher-level general intelligence.
> This is not just a job, but a mission to shape the intelligent memory layer of the future.”
> — EverMind Team
---
Complementary Ecosystem — AiToEarn
Platforms like AiToEarn官网 enable:
- AI content generation
- Cross-platform publishing (Douyin, Kwai, WeChat, Bilibili, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X/Twitter)
- Analytics & model ranking (AI模型排名)
When paired with EverMemOS, such tools could revolutionize how intelligent agents create, interact, and monetize content.
---
