Memory

In Module 4, you wrote a system prompt. That prompt defines how your agent works. It is the foundation — and it matters enormously.

But here is what separates the people earning consistently from the people who plateau after a few weeks: memory. Specifically, the accumulated knowledge, examples, and lessons that your agent can draw on beyond its built-in training.

An agent with a great system prompt and no memory produces good work from day one. An agent with a great system prompt and rich memory produces work that gets noticeably better every month — because it is drawing on a growing library of examples, client preferences, proven patterns, and hard-won lessons.

Memory is what turns your agent from a capable worker into a compounding business asset. And the best part: building it does not require technical skill. It requires attention, five minutes after each significant job, and a monthly maintenance habit.

Do not start by trying to build perfect, comprehensive memory on day one. You will build memory in layers over your first month, then continue adding to it monthly.

You need exactly two things before your agent handles any live work.

Two to three examples of excellent output are the single most impactful memory documents you will create. These are examples that go in the system prompt's Examples section. But they are also memory. They teach the agent what good looks like in a way no written rule can fully capture. If you have past work: choose your two or three best pieces. The ones you would confidently show any client as proof of quality. If you are starting fresh: create two examples manually. Write them yourself, or find excellent examples from other sources and adapt them to your service. Annotate them — explain what makes each one strong.

A process template is the step-by-step description of how your agent handles a standard task. You already built this into your Instructions section in Module 4. As a separate memory document, it can be more detailed — including tips, common pitfalls, and nuances that do not fit in the system prompt. For a writing agent: Opening paragraph structure should start with the most interesting observation, then a specific detail that supports it, then a bridge sentence into the article body. Never open with a question. Never open with a definition. For a research agent: Source priority should be official company website first, LinkedIn second, industry publications third, general web last. Always verify the company is still active before including it.

With just these two pieces — examples and a process template — your output quality will be significantly better than instructions alone.

Once your agent has handled some real work, you have the experience to build meaningful reference documents.

A style guide defines your tone, vocabulary preferences, structural conventions, and things to avoid. Even a single page makes a noticeable difference. Key sections include: Voice and Tone (describe your voice in 2-3 adjectives; specify person such as second person; clarify contractions). Vocabulary (words you prefer such as clear, specific, concrete, actionable; words to avoid such as leverage, synergies, holistic, robust, seamlessly, delve; guidance on industry terms). Sentence Structure (short sentences for emphasis; longer sentences for explanation; one idea per sentence; one theme per paragraph). Openings (what works like leading with a surprising fact or bold statement; what to avoid like opening with in this article). Closings (how to end such as with one clear, actionable recommendation). Formatting Conventions (heading hierarchy like H2 for main sections, H3 for subsections; emphasis rules for bold and italics).

Quality standards define what separates excellent output from acceptable output in your service. What does a five-star piece look like versus a three-star piece? Write it down explicitly so the agent has a benchmark beyond your examples.

Now you have real experience to draw from. This is where episodic memory — the record of what has happened — starts building.

The experience log is a record of significant jobs. After each significant job, spend five minutes writing a brief note. Three types of notes work best: A pattern note captures what worked, such as 2026-02, B2B SaaS explainer. The client specifically responded to the use of analogies for technical concepts. Three analogies in a 700-word piece hit the right balance. Apply analogies to all technical explainer content. A lesson note captures what you would do differently, such as 2026-02, Research brief with geographic filter. Two records slipped through from wrong regions because the country filter was ambiguous. Always confirm the specific countries/regions at intake for geographically targeted research. An edge case note captures how you handled something unusual, such as 2026-02, Client asked for formal AND casual versions of same piece. Delivered casual version first with a note explaining the approach, offered formal as alternative. Client chose casual, rated 5 stars. This approach works for dual-version requests.

Memory is a living system. It needs maintenance.

New examples should always represent your current quality level, not where you started. When you produce work you are particularly proud of, or when a client gives exceptional praise for a specific piece, add it to your example library.

Refined process templates improve through real work. As you discover process improvements through real work, update your templates. The process template you use in month six should be sharper and more detailed than the one you started with.

Lessons from revisions teach you what your agent needs to learn. Every revision request is data. What did the client ask for? Was it something your agent could have anticipated? Could a rule or a better example prevent this revision in the future? If yes, add the learning.

Domain knowledge fills specific gaps in your agent's output. As you identify specific knowledge gaps — terminology it gets wrong, facts it should know, conventions in your industry — create or update domain knowledge documents to fill those gaps.

Even excellent content is hard for an agent to use if it is poorly organised.

Use clear headings. The agent navigates your documents to find relevant information. Headings make information findable. Unlabelled prose makes everything harder.

Put the most important information first. Universally applicable standards and rules go at the top. Situational or context-specific information goes lower. If the agent only reads the first section (which sometimes happens with large documents), make sure the first section contains the essentials.

Separate documents for separate functions. Do not put your style guide, process template, and experience log in the same file. Keep them as clearly labelled separate documents. Clear separation means the agent (and you) can find what is needed quickly.

Use structured formats for structured information. Rules belong in lists, not paragraphs. Examples belong in consistent input/output pairs. Procedures belong in numbered steps. Match the format to the content.

Review and trim regularly. Memory that grows without pruning becomes cluttered. Outdated client notes, superseded examples, and lessons that have already been incorporated into your rules are noise. Remove them.