Haiyun Hu

I'm Haiyun, founder of HPSILab. If you're building an AI workflow for stock research, here's a compact acceptance test you can adapt to any data provider.

Scenario: a model's directional signal is positive, while volatility is elevated and an options pressure level is nearby. These inputs describe different things. The agent should preserve the disagreement rather than turn them into one confident recommendation. This is an illustrative scenario, not a current market observation.

Ask your agent to produce five fields:

1. Evidence: each input, its source, and the available as-of date.

Most of us now use 2 4 AI tools, and every one of them starts from zero. We put together an honest comparison of the 5 common ways to build an AI second brain :

1. Markdown files: free and portable, but lots of copy-paste

2. Local files (CLAUDE.md, rules files): best for coding agents

3. ChatGPT / Claude / Gemini Projects: easiest if you use one assistant

4. Drive / GitHub connectors: good for large archives

5. A portable memory layer over MCP, which is what we build with AI Context Flow

Projects is enough if you only use one assistant. Whereas a memory layer helps iwhen you switch between models, share context with a team, or run workflows (n8n, agent teams) that need to read and write the same memory.

AI agents are getting pretty good at finding information, summarizing documents, and answering questions. But what about letting them change the information itself?

Imagine an AI agent finds outdated onboarding instructions in your company's wiki. Should it be allowed to fix them immediately? Suggest an edit for someone to review? Or simply flag the problem and leave everything to a human?

We've been thinking about this while developing the XWiki MCP Server.

Our approach is to respect existing user permissions and keep write operations disabled by default. Administrators can enable them when needed.