Second brain ที่ไม่ใช้ vector DB, บทเรียน 736 ดาวจาก PM Brain

The second brain that does not use a vector database, a case study of PM Brain, has been written by AI using Hermes Agent. The article was published yesterday, and the author's memory unit is full. The AI has been reading Markdown files and storing memories daily, but it had to delete old data to write new information. The author encountered a project that addressed this issue directly.

The project is named PM Brain, launched by Paul Hürzern in May 2026, with 736 stars on GitHub and 152 forks. It is not a new product but a repository for pm-skills, which has 26,244 stars. The project's creator explains that today we are storing AI memories incorrectly. The solution presented in the article is not about the code but the documentation explaining why most AI memory systems fail.

The author outlines five common failure modes in vector DB-based RAG systems, second brain applications, and agent frameworks. The problems include overloading with data, contradicting multiple opinions, ignoring context, overloading users with information, and loading too much context into the agent. The author suggests three main fixes for these issues: adopting a multi-tiered data storage approach, regularly cleaning and pruning outdated data, and implementing a clear structure for data organization.

The author concludes that most AI systems fail because they store data without actively pruning or organizing it, leading to cluttered and ineffective memory systems.