The augmented LLM

Before any clever architecture, there’s this: a bare LLM can only turn text into text. It can’t read your files, search the web or remember yesterday. The augmented LLM is the model plus its equipment — and it’s the LEGO brick every other pattern is built from.

The analogy

A brilliant consultant locked in a room with no phone, no computer, no archive: that’s a bare LLM. Now give them a phone line (tools), a library card (retrieval) and a notebook (memory) — same brain, but now they can actually work. That’s the augmented LLM.

The principle

flowchart LR
    IN([input]) --> LLM[LLM]
    LLM --> OUT([output])
    LLM <--> R["retrieval — know"]
    LLM <--> T["tools — act"]
    LLM <--> M["memory — remember"]

Three augmentations, each answering one weakness:

Retrieval — fetch relevant documents into the context (see RAG), so answers come from your data.
Tools — let the model trigger actions via tool calls: query a database, call an API, run code. MCP is the standard socket for this.
Memory — persist useful facts across turns or sessions, instead of relying on the context window alone.

A concrete example

A support assistant that actually helps:

User: "Where is my order #4521?"
LLM → tool call: get_order(4521)        # tools
LLM ← {status: "shipped", eta: "Jul 8"}
LLM → retrieval: shipping policy         # retrieval
LLM: "It shipped and arrives July 8.
      Under our policy you can still…"

Without augmentation, the model could only guess. With it, every claim is anchored in a real lookup.

When to use it

Always. It’s not an option, it’s the foundation — every pattern that follows assumes an augmented LLM.
Start by asking: what does my model need to see (retrieval), do (tools) and remember (memory) for this task? Add only that. Every tool schema costs input tokens on every call.

The classic trap

Over-equipping. Twenty tools “just in case” clutter the context and confuse the model’s choices. The best augmented LLM has the fewest tools that cover the job — like a good consultant with one phone, not six.