Notes on building MemoryFirst, the AI-native memory layer for businesses. Engineering, product, and the occasional founder rant.
Karpathy framed LLMs as a new operating system: the model is the CPU, context is the RAM. But no serious computer ships without a disk. Persistent memory is the missing layer.
Jun 18, 2026
Mem0, Supermemory, Zep, MemoryFirst. They all promise 'memory for AI'. The differences that actually matter: where your data lives, whether you can audit answers, and whether you can self-host.
Jun 18, 2026
An AI that asserts things it can't prove is a liability, not a feature. Here's how a memory layer with citations, retraction and EU data turns trust into something verifiable.
Jun 18, 2026
A memory layer is only useful if your tools can read it. That's why MemoryFirst speaks two protocols you already use: MCP and the OpenAI-compatible chat API. No proprietary SDK.
Jun 18, 2026
A good answer starts with good retrieval. Here's how we combine semantic, lexical and graph search, fuse them with RRF and rerank with a cross-encoder — and why each piece matters.
Jun 18, 2026
For many teams, a memory layer in someone else's cloud isn't an option. That's why MemoryFirst is built to run on your own infrastructure, in the EU, with no phone-home.
Jun 18, 2026
Why we're building the memory layer for AI-native businesses, why now, and what's coming next.
2026-05-20