ReFlex.AI: Building a Persistent Cognitive Architecture for Long-Running AI Agents

ReFlex.AI is an open-source research project dedicated to solving the problems of memory degradation, identity drift, and hallucinations in long-running AI agents. Through a layered memory system and a self-correcting cognitive loop, it provides LLM agents with persistent state management capabilities. The project adopts an ROCm-first strategy, supports AMD hardware, has open and reproducible code, and aims to become a reliable infrastructure for long-running AI applications.

Current LLM agents rely on volatile context windows, leading to five core issues:

1. Context fragmentation: Historical records are lost when sliding out of the window, reducing conversation quality;
2. Memory degradation: Repeated summarization causes information distortion;
3. Identity drift: Without persistent anchoring, goals and personality traits shift;
4. Historical fabrication: Making up unoccurred events;
5. Unreliable long-range reasoning: Logical consistency decays with conversation length. These are default problems when stateless models exhibit stateful behavior.

ReFlex.AI draws inspiration from biological cognition and adopts three core design principles:

1. Layered memory subsystem: Similar to computer cache hierarchy, information is promoted/demoted/compressed between layers;
2. Cognitive loop: A closed loop of execution → observation → reflection → correction → memory writing;
3. Authenticity reconciliation: A consistency layer checks for factual drift and fabricated memories. The layered memory system includes five levels:

• Short-term buffer: Minute-level volatile storage for recent interactions;
• Working memory: Volatile storage for current tasks, bound to the context window;
• Episodic memory: Session-to-day persistent storage with timestamped event records;
• Semantic memory: Long-term persistent storage that extracts facts/entities/relationships;
• Compressed archive: Cold storage for over months, summarizing long-tail history. Information flow follows rules of promotion, demotion, and compression to balance resource usage and history management.

The core innovation is the closed-loop self-correcting cognitive loop:

1. Execute actions and respond;
2. Observe results;
3. The reflection engine evaluates consistency (goal achievement, unexpected events, experiential learning) and writes to episodic memory;
4. Consistency protection layer checks: factual drift, fabricated memories, invalid reasoning, output contradictions;
5. After correction, write to memory and return to the execution phase. This loop allows agents to continuously improve and avoid repeating mistakes.

Adopts an ROCm-first strategy and supports AMD hardware:

• Hardware: AMD Instinct MI300X/MI325X/MI350X series, with planned support for MI400;
• Compute stack: ROCm7.x (HIP/RCCL, etc.), ROCm version of PyTorch;
• Inference services: vLLM ROCm version, SGLang;
• Training and fine-tuning: Hugging Face + Optimum-AMD/PEFT/LoRA;
• Storage and retrieval: FAISS/pgvector vector retrieval, SQLite/PostgreSQL persistence;
• Runtime: Python3.11+ asynchronous architecture, custom test framework. Provides an alternative to NVIDIA solutions.

Applicable to long-running AI applications:

1. Personal AI assistants: Remember preferences, conversation history, and long-term goals;
2. Enterprise knowledge management: Continuously learn company history and culture, answer context-aware questions;
3. Automated workflows: Long-term tracking of complex tasks (e.g., project management);
4. Research analysis: Continuously track literature/experiments and maintain knowledge graphs.

ReFlex.AI redefines AI agent architecture by taking memory as a core design element:

• Engineering path: Directly solve the amnesia problem instead of covering it up;
• Open-source contribution: Release reproducible research and infrastructure;
• AMD ecosystem: Provide a feasible solution for non-NVIDIA deployments;
• Future: Promote a more reliable and coherent AI assistant ecosystem.

For developers and researchers:

• Closely follow the development of the ReFlex.AI project and use its open-source resources to build long-running AI applications;
• Explore the application of layered memory and self-correction mechanisms in real-world scenarios;
• Try ROCm-based hardware deployment to reduce ecosystem lock-in risks.