AI and Knowledge: does it learn? How can I give it my entire project as context?

I believe this is a concept we all must clearly understand: Large Language Models (LLMs) do not inherently have the ability to “know your context” or “have your full project” stored internally. Instead, they’re using specific techniques to enrich the queries you send them.


When you train an LLM, you generate internal weights that allow the model to think and respond in specific ways. Fine-tuning adjusts these weights further to alter behavior. This results in implicit knowledge: during training, the LLM condenses statistical patterns from vast datasets into neural weights, enabling it to produce relevant facts or answers. However, this implicit knowledge is static, it doesn’t update automatically after training.

The capacity to “learn” dynamically from new, changing data is handled externally. That additional context is stored separately and passed along with every query as extra context. Another layer of software prepares the prompts by injecting this relevant contextual information before it reaches the LLM.

A clear example of this approach, which you can even explore by checking out its open-source code, is Continue.dev. Continue automatically assembles context from your current file, recent conversations, and related past information retrieved using semantic search, powered by a vector database (ChromaDB). Semantic search allows Continue.dev to quickly recall relevant previous interactions or files based on meaning rather than exact wording. It then sends this assembled context to the LLM, ensuring precise, relevant, and useful responses.

The actual message sent to the LLM is a prompt clearly explaining the context and purpose in plain text. The LLM then processes this exactly as if you had manually copied and pasted all relevant information. If you use Ollama integrated with Continue.dev, you can directly see these prompts being sent and verify what I’m explaining here.

It’s not as simple as merely setting up a vector database, querying it, and forwarding results to the LLM. There are numerous advanced techniques for context splitting, RAG, agentic RAG, and more. However, the fundamental principle remains unchanged.

At DisplayNote, we’re embracing this principle.

We’re actively working with AI tools, not just exploring their capabilities, but also applying best practices around prompt design, context management, and real-world testing. Our goal is to understand where these technologies add genuine value, and to integrate them thoughtfully into our workflows and products. By experimenting, refining, and sharing what we learn, we’re helping shape an approach to AI that’s practical, responsible, and grounded in real outcomes.