Start Here: From Memory to Agents

Welcome. Before anything technical, hold onto one idea: working well with AI is not about finding a smarter model or a slicker demo. It is about the judgment you bring to the work — what you are really trying to achieve, what must stay human, and what can safely be handed to a system that runs on its own.

That matters more every month, because these tools have crossed a line: they no longer just answer — they remember, retrieve, decide, and act. And a persistent agent amplifies whatever structure you give it. Give it a vague goal and it scales the vagueness; give it a clear one, the right context, and sane boundaries, and it compounds your judgment instead of your mistakes.

So this course has one job: to help you give these systems good structure — and it moves in two parts. First, these seven foundation lessons build the shared vocabulary the rest of the course assumes, tracing how working with language models evolved: from a model that simply remembers your preferences, to systems that retrieve knowledge, manage their own context, reach into your tools, and finally act on their own. Each stage did not replace the one before it — it absorbed it. A modern agent still uses a system prompt, still depends on good prompting, still retrieves with RAG, still lives or dies by context engineering, and still reaches the world through protocols like MCP. Then the fourteen lessons turn that vocabulary into practice, on real problems you care about.

Read this first part as a layered stack — every layer adds capability, and a new kind of responsibility. The diagram below is the whole territory at a glance.

We begin where every layer begins: with what the model already knows, and what you choose to tell it.

The simplest way a model is shaped is by what sits in front of every request. Two mechanisms do most of the work: the system prompt and memory / preferences.

The system prompt

A system prompt is a standing instruction prepended to every conversation. It sets the model's role, tone, rules, and boundaries before the user says anything. Where a user message asks for one thing, the system prompt governs how the model behaves across all things — "You are a careful financial assistant; never give tax advice; cite sources." It is the cheapest, highest-leverage control surface you have, and it is the seed from which every later technique grows.

Memory and preferences

Early assistants were stateless: each request started from zero. Memory changed that. Consumer products now persist facts about you ("I write in British English", "I run a nonprofit") and silently inject them into context on future turns. Persistent agent runtimes take this furthest: Hermes Agent (Nous Research) ships a built-in learning loop that creates skills from experience and builds a deepening model of who you are across sessions, and OpenClaw maintains long-term memory of your preferences across every channel you message it on. Research frames this as the defining feature of modern agents — a memory mechanism that lets a system accumulate and reuse experience rather than re-deriving it every time.

Example. You tell an assistant once that your company is called Centpol and you prefer concise bullet answers. A stateless model forgets immediately. A memory-enabled one stores both facts and applies them next week without being reminded — the first, faint version of an agent that "knows" your world.

The trade-off arrives immediately: memory that helps can also leak, drift, or go stale. Deciding what deserves to be remembered is the first act of context discipline — the through-line of this entire course.

References & further reading

• Nous Research (2026). Hermes Agent — Documentation. hermes-agent.nousresearch.com.
• Steinberger, P. (2026). OpenClaw — Personal AI Assistant. openclaw.ai.
• Brown et al. (2020). Language Models are Few-Shot Learners (GPT-3). NeurIPS.
• Zhang et al. (2024). A Survey on the Memory Mechanism of LLM-based Agents. arXiv.
• Anthropic (2024). Prompt Engineering Overview. Claude Docs.