Your context is your moat

Building persistent memory systems for LLM workflows taught me something counterintuitive: the model matters less than the context you feed it.

I spent months building Recall — a shared, attributed memory system for AI coding agents. The thesis was simple: if agents could remember what happened across sessions, across team members, and across projects, they'd stop being stateless tools and start being actual collaborators.

What I discovered

The gap between a "dumb" model with great context and a "smart" model with no context is enormous — and it goes the wrong way from what most people expect. A smaller model that knows your codebase architecture, your team's conventions, your past decisions and their rationale, and your current priorities will outperform a frontier model starting cold every single time.

This has implications for how we should think about AI tooling:

• Memory is infrastructure, not a feature. It's not something you bolt on — it's the foundation everything else builds on.
• Attribution matters. When a team shares a memory graph, knowing who contributed each piece of knowledge is as important as the knowledge itself. "Sarah said this API is flaky" carries different weight than an anonymous note.
• Forgetting is as important as remembering. A memory system that never prunes becomes noise. The hard engineering problem isn't storage — it's relevance.

The moat

Every team that builds up a rich, well-attributed context layer creates something no competitor can replicate by switching models. Your institutional knowledge — the decisions, the gotchas, the "we tried that and here's why it didn't work" — is genuinely unique. It's the one thing that can't be commoditized.

Models will keep getting better and cheaper. Context — real, accumulated, team-specific context — only gets more valuable. That's your moat.