This lesson is being recorded

Every video of this course is being uploaded to my YouTube channel. Subscribe so you do not miss it.

Bruno AshfordBruno Ashford @brunoashfordSubscribe

Module 4 · Lesson 4.1 · 10 min

Why AI Forgets

Understand the context window as working memory with a size limit, and why a written handbook fixes most forgetting.

Module 4
Your Business Brain

Write your business down once. Every session starts knowing it.

Bruno Ashford
Bruno Ashford · Module 401 / 12
The problem
On Monday it wrote in your voice. On Tuesday it had never heard of you.
Bruno Ashford
Bruno Ashford · Module 402 / 12
Why AI forgets

The context window is a desk

If it is not on the desk, Claude does not know it.

  • Every session starts with a clear desk
  • Files, messages and replies fill it
  • When it is full, quality drops
  • /context shows what is on it
Bruno Ashford · Module 403 / 12
Two kinds of memory

What carries over between sessions

CLAUDE.mdAuto memory
Who writes itYouClaude
What it holdsInstructions and factsLearnings and preferences
LoadedEvery sessionEvery session
Edit with/memory or any editor/memory
Bruno Ashford · Module 404 / 12
The handbook

What goes in your CLAUDE.md

  1. 1CompanyWho we are in one paragraph
  2. 2OffersWhat we sell and for how much
  3. 3ClientsWho we serve and who we do not
  4. 4VoiceConcrete rules, real examples
  5. 5Red linesWhat never happens
  6. 6Folder mapWhere everything lives
Bruno Ashford · Module 405 / 12
Guidance from Anthropic's docs

Keep the handbook lean

~200lines per CLAUDE.md, recommended ceiling
1rule per line
@imports for longer documents
Bruno Ashford · Module 406 / 12
Rules

Red lines, standards, quality bars

Three different jobs. Write them separately.

  • Red lines: must never happen
  • Standards: how we always do it
  • Quality bars: checked before you see it
  • Test every red line by breaking it
Bruno Ashford · Module 407 / 12
Guidance vs. enforcement
CLAUDE.md guides the agent. Permissions and hooks lock the door.
Bruno Ashford
Bruno Ashford · Module 408 / 12
Folders

If a new hire could find it, so can the agent

Structure is context you never have to type.

  • One folder per business area
  • Dates first: 2026-09-proposal.md
  • Originals never edited
  • A folder map in CLAUDE.md
Bruno Ashford · Module 409 / 12
Long sessions

Four commands that keep answers sharp

  1. 1/contextSee how full the desk is
  2. 2/compactSummarise and keep going
  3. 3/clearFresh start between tasks
  4. 4Handover noteSave decisions to a file
Bruno Ashford · Module 410 / 12
Your project

A tested company handbook

Write it, then prove it on three real tasks.

  • Loaded and visible in /context
  • Knows your client, offer and voice
  • Refuses two red-line tests
  • Under about 200 lines
Bruno Ashford · Module 411 / 12
Recap
Stop re-explaining your business. Write it down once, and let every session start from there.
Bruno Ashford
Bruno Ashford · Module 412 / 12
01 / 12

Use ← → to move. The marker highlights text and the pencil draws on the slide; navigation locks while you annotate. Undo removes the last mark, Esc exits.

1Think of the context window as a desk

Everything Claude is working with sits on a desk: your messages, the files it opened, its own replies, your handbook. The desk is large but not infinite. When it fills up, older things get summarised or pushed off, and quality drops.

2See what is on the desk right now

Type /context in any session. Claude Code shows how full the context is and what is taking up space, including which memory files were loaded. Do this once now so the idea stops being abstract.

Terminal
/context
3Know the two kinds of memory

Claude Code carries knowledge between sessions in two ways. CLAUDE.md files are instructions you write, loaded at the start of every session. Auto memory is notes Claude writes for itself when you correct it or state a preference. Both are loaded as context, and you can view and edit both with /memory.

Terminal
/memory
4Test the forgetting yourself

Start a fresh session and ask Claude something only your team would know. It will guess or admit it does not know. Keep the question: after lesson 4.2 you will ask it again and see the difference.

Prompt for Claude
Without looking at any files, tell me: who is our ideal client, what is our main offer, and how do we sound in emails?

If you have used AI for more than a week, you have felt this. On Monday it wrote a perfect proposal in your voice. On Tuesday it had no idea who your clients were and wrote like a bank. You did not change anything. It just forgot.

It did not forget out of carelessness. Every session starts with a fresh working memory called the context window. Whatever is not in that window does not exist for Claude. The fix is not to type more each morning. The fix is to write your business down once, in a file Claude reads automatically every time it starts.

Key points

  • Every session starts with a fresh working memory.
  • If it is not in the context window, Claude does not know it.
  • /context shows what is taking up space.
  • CLAUDE.md is what you write. Auto memory is what Claude notes.
  • Write your business down once instead of retyping it daily.

Avoid

  • Blaming the model when the real problem is missing context.
  • Pasting the same background into every conversation.
  • Assuming a long chat history means Claude remembers everything in it.
  • Never checking what is actually loaded with /context.
Real-world example: an accounting practice

A four-partner accounting practice was spending around 10 minutes at the start of every AI session pasting in the same background: services, fee structure, the rule that tax advice always goes through a partner. Across six staff that was roughly five hours a week of copy and paste. Once that background lived in a handbook file loaded automatically, the pasting stopped and the answers became consistent across the whole team.

Bruno Ashford Subscribe