Memory and RAG: the next step towards source-grounded work

This is a development direction, not an announcement of released RAG. The reviewed NC AgentOS version has file memory and text import; semantic RAG retrieval is disabled.

The intended workflow

A user selects documents, sees processing results and connects a knowledge base to an agent. Answers should show retrieved passages and precise source references. The system should say when it has no supporting evidence.

What must be validated

• Format parsing and document update/deletion without orphaned passages.
• Retrieval quality against controlled questions and quotations.
• Isolation between agents, projects and users.
• Access rights and defence against instructions hidden in source documents.
• Clear errors for users without technical experience.

Improvement through evaluation

Feedback, verified solutions and repeated tests can support a managed improvement process. A full self-improvement cycle still requires implementation and measured evaluation. This is not a promise of automatic training of model weights.

Describe your sources and expected answers to help shape development priorities.