Explainer

What Is an AI Harness?

An AI harness is the software built around a language model that gives it memory, structure, and the ability to act over time — as opposed to the model itself, which by default only responds to whatever is in its current context window and forgets everything once the conversation ends.

Every frontier model — Claude, GPT, Gemini — is extraordinarily capable at the moment you're talking to it and has no memory of you the moment you close the tab. That's not a bug; it's how the underlying technology works. A harness is what fixes the gap: it stores what the model would otherwise forget, decides which model (or which size of model) should handle a given task, and gives the whole system a reason to do things without being asked again each time.

The chatbot vs. the harness

A chatbot is a thin wrapper over a model: you type, it responds, and the conversation is the entire product. Useful, but the burden of remembering context, restating goals, and deciding what to do next always falls back on you. A harness inverts that. It's the layer of memory, orchestration, and follow-through wrapped around the model — the difference between an assistant you have to drive every single turn and one that can actually manage something on its own.

Where Refractive fits

Refractive is built as a harness, not a model — explicitly model-agnostic, running by default on open-weight models it pays for (DeepSeek V4 Pro, GLM-5.2, or Kimi K2.6), or on whichever of Anthropic, OpenAI, Google, or OpenRouter's models a user connects instead. What Refractive itself is responsible for is everything a raw model doesn't do: a living map of what it knows about a person's life (the Context Graph), the judgment to route different kinds of work to different models based on cost and complexity, and the follow-through to turn a conversation into a tracked project instead of a transcript that evaporates. The model does the reasoning in the moment; the harness is what makes that reasoning cumulative.

Why the distinction matters

If you've ever re-explained your situation to a chatbot for the third time this week, you've felt the absence of a harness firsthand. The distinction isn't academic — it's the difference between a tool you have to actively operate and one that can carry context forward well enough to feel like it actually knows you.