I expected maybe 20 people to see it. Instead it got: Over 10M views 10k stars on GitHub And thousands of other developers building with it.

Over 1,000 comments came in, some were very harsh, but I also got some genuine feedback. I spoke with many of you and spent the last few months building Memvid v2: it’s faster, smarter, and powerful enough to replace entire RAG stacks.

Thanks for all the support.

Ps: I added a little surprise at the end for developers and OSS builders

TL;DR - Memvid replaces RAG + vector DBs entirely with a single portable memory file. - Stores knowledge as Smart Frames (content + embedding + time + relationships) - 5 minute setup and zero infrastructure. - Hybrid search with sub-5ms retrieval - Fully portable and open Source

Try it here: Memvid.com

—----------------------------------------------------------------

Some background: - AI memory has been duct-taped together for too long. - RAG pipelines keep getting more complex, vector DBs keep getting heavier, and agents still forget everything unless you babysit them. - So we built a completely different memory system that replaces RAG and vector databases entirely.

What is Memvid: - Memvid stores everything your agent knows inside a single portable file, that your code can read, append to, and update across interactions. - Each fact, action and interaction is stored as a self‑contained “Smart Frame” containing the original content, its vector embedding, a timestamp and any relevant relationships. - This allows Memvid to unify long-term memory and external information retrieval into a single system, enabling deeper, context-aware intelligence across sessions, without juggling multiple dependencies. - So when the agent receives a query, Memvid simply activates only the relevant frames, by meaning, keyword, time, or context, and reconstructs the answer instantly. - The result is a small, model-agnostic memory file your agent can carry anywhere.

What this means for developers:

Memvid replaces your entire RAG stack.

- Ingest any data type - Zero preprocessing required - Millisecond retrieval - Self-learning through interaction - Saves 20+ hours per week - Cut infrastructure costs by 90% - Just plug Memvid into your agent and you instantly get a fully functional, persistent memory layer right out of the box.

Performance & Compatibility (tested on my Mac M4) - Ingestion speed: 157 docs/sec - Search Latency: <17ms retrieval for 50,000 documents - Retrieval Accuracy: beating leading RAG pipelines by over 60% - Compression: up to 15× smaller storage footprint - Storage efficiency: store 50,000 docs in a ~200 MB file - Memvid works with every model and major framework: GPT, Claude, Gemini, Llama, LangChain, Autogen and custom-built stacks.

You can also 1-click integrate with your favorite IDE (eg. VS Code, Cursor)

If your AI agent can read a file or call a function, it can now remember forever.

And your memory is 100% portable: Build with GPT → run on Claude → move to Llama. The memory stays identical.

All 100% open-source and available today.

Memvid V2 is the version that finally feels like what AI memory should’ve been all along.

If this sounds useful for what you’re building, I’d love for you to try it and let me know how we can improve it.

GitHub: https://github.com/memvid/memvid Website: https://memvid.com