Anthropic shipped a source map inside an npm package. The entire Claude Code codebase.
Buried in that code is AutoDream: a background engine that consolidates, prunes, and reorganizes context while the user is idle. Four layers of structured memory. A process that runs while you sleep, merging observations, killing contradictions, converting vague notes into durable facts.
They built a context engine.
We're building a company on the thesis that structured context is the only durable advantage in AI. So — did Anthropic just make us irrelevant?
What's in the code
Link. Paste it into your LLM and ask questions.
Claude Code has four memory layers: user-written instructions, agent-generated notes, session context, and AutoDream for consolidation between sessions. (They are so good with names — ours is just Eval.)
This is serious infrastructure for persistent, structured context.
Their engineers built this because raw model intelligence isn't enough. Context degrades. Information contradicts itself. The hundredth session is worse than the first unless you actively fight entropy.
Different layers
AutoDream optimizes the working memory of one tool on one project. That's micro — session notes, observed patterns, accumulated knowledge about a codebase.
The context we're trying to structure sits a level above. The operational knowledge of a company — how it communicates, what it can and can't say, who it talks to, what worked before. That context exists before any AI tool is opened.
What's interesting is that what they built and what we're building look structurally similar. They did it for a coding agent. We're trying to do it for every agentic system that produces content. That convergence is probably what the future looks like in some way.
Where I actually am
What I know. Context management is the central infrastructure challenge of AI. The biggest lab in the world just proved it.
What I don't know. Whether the solution lives inside the platforms or outside them.
What I think. Context is inherently cross-platform. Your brand intelligence shouldn't be locked inside one tool's memory. It shouldn't restart from zero when you switch models.
That's the bet behind .fylle — an open protocol for portable access to context. Whether that bet is right is something I can't prove yet.