Build the Memory Surfaces

Lesson outcome

You will build the minimum memory system Claude needs in order to stop repeating the same mistakes and start working from stable context.

Why this matters in an agency

Without durable memory, the agency pays a hidden tax every time it restates the same preferences, definitions, and process rules. That tax shows up as longer prompts, weaker continuity, and lower trust in the model. Memory surfaces reduce that tax by moving recurring context out of the chat and into documents Claude can reference again and again.

Inputs, tools, and prerequisites

You need a root instruction file, a lessons log, and a small set of working notes for active projects. If you already use AGENTS.md or CLAUDE.md in repos, that is a good starting point. If not, create them now as lightweight operating documents. Your vault should also have a place for business-level notes that outlive any single project.

Step-by-step walkthrough

Start with CLAUDE.md or the equivalent root instruction file. This document should explain how Claude is expected to operate in the environment: tone, guardrails, verification standards, file-editing constraints, and any recurring project-specific rules. Keep it short and practical. You are not writing a manifesto. You are writing the instruction block that prevents repeated confusion.

Next create a lessons log. This file exists to capture corrections, decisions, and patterns discovered during work. Every time a mistake happens or a process improves, write the smallest useful lesson that would help the next run. Good lessons are explicit and preventative. Example: "When updating DAGA offer structure, cascade changes across vault docs, app copy, and legal surfaces." That kind of note compounds because it prevents recurring drift.

Then create working notes for active initiatives. These are not permanent strategy notes. They are operational notes that help Claude orient quickly: what the task is, what has already been decided, what is blocked, and what success looks like. If the initiative finishes, either archive the note or promote any durable insight into a permanent knowledge note.

Finally decide what belongs in memory and what belongs in a one-time prompt. Durable memory should include instructions, naming conventions, architecture facts, offer definitions, stable process rules, and durable lessons. One-time prompts should include the immediate task, current context, and temporary constraints. If you mix the two, memory becomes noisy.

Failure modes and verification checks

The main failure mode is turning memory into a junk drawer. Another is using memory files to store transient task chatter that should expire. Verification is straightforward: can Claude start a new session and quickly understand the environment, the operating rules, and the most important lessons without you restating them? If yes, the memory surfaces are working.

Implementation checklist

• Create or refine CLAUDE.md.
• Create a lessons log.
• Create working-note templates for active initiatives.
• Separate durable instructions from temporary task context.
• Review memory notes monthly for noise and drift.

Immediate next action

Write the first three entries in your lessons log from recent mistakes or repeated clarifications in the business.