Chitta: Giving Self-Hosted Models Your Business Memory Without Retraining

The Retraining Trap

For a long time, the answer to "make the model understand our business" was to retrain it on your data. That approach is expensive, slow, and stale the moment your projects, people, and decisions change. Worse, it bakes sensitive information into model weights, where it is nearly impossible to update or remove.

There is a better way. Instead of changing the model, give it memory.

Memory, Not Retraining

Chitta is the memory layer built into the 100X Code CLI. It maintains a permission-aware knowledge graph and vector store of your organization - your projects, people, decisions, and documents - and feeds the right context to the model at the right moment. The model stays general; the memory makes it yours.

Because the context lives outside the weights, it updates instantly, respects access controls, and can be inspected or deleted at any time. No retraining runs, no stale snapshots, no sensitive data fused into a checkpoint.

Why It Pairs With Self-Hosting

Persistent memory of your business is only safe if it stays inside your perimeter. Chitta runs entirely within your environment, alongside your private LLM deployment - the knowledge graph never leaves the building. It is one component of the harness layer that turns an open model into a system that actually knows your organization.

What It Looks Like in Practice

Ask our coding assistant to build a component, and it applies the conventions your team already agreed on - named exports, your styling system, your directory structure - because it remembers them from earlier sessions. Across projects, the same preferences carry over. The AI is not smarter than an open model; it simply knows more about you, and it learned that without a single retraining run.