Why I stopped using Vector DBs for AI Memory (and used Git + Markdown instead)

#architecture #ai #opensource #git

Hey DEV community! 👋

Like many of you, I've been heavily experimenting with AI Agents recently. But as a software architect who hates unnecessary overhead, watching the community default to spinning up heavy Vector DBs and complex graph frameworks just to store conversation context felt like massive over-engineering.

We often use massive systems to solve small problems. I wanted an AI memory system that is zero-cost, serverless, and completely transparent. So, I built a lean alternative.

🔥 Talk is cheap, here is the repo: https://github.com/phucphungbk/lean-ai-memory

The Lean Approach: Back to Basics

Instead of adding new layers of infrastructure, I looked at what we already have:

Git as the Core Engine: We already use Git for version control. It turns out it's absolutely perfect for managing conversation history. You can easily track, diff, and rollback an AI's "thought process" just like reverting a bad commit.
Markdown as the Storage Format: It’s lightweight, humans can read and debug it instantly, and LLMs parse it perfectly without needing complex embedding pipelines.
Zero-cost & Portable: It can be packaged as an independent module and integrated directly into internal automation tools without incurring any DB maintenance costs.

Why I open-sourced this

I’m sharing this with a completely open mindset. Instead of optimizing it in a silo, I want to see how this framework holds up in the wild. I'm highly anticipating the community bringing their own battle-tested custom rules into the system to push its boundaries.

I’d love for you guys to clone it, tear the architecture apart, run some tests, and drop your feedback or PRs.

Let me know what you think in the comments! Happy coding! 🚀