Phase 1, Session 4

Why Claude "Forgets" — Not a Bug, an Architecture

The Big Idea

When Claude seems to forget something you told it, people assume it's a mistake or a limitation in the AI. Sometimes it is. But most of the time, it's one of three very specific, very different architectural realities — each with a different cause and a different fix.

Lumping all "forgetting" together is like saying your car "doesn't work" whether it has a flat tire, an empty gas tank, or a dead battery. True in all three cases. Completely different solutions.

Analogy
Claude has no long-term memory at all. It's not like a person who forgets over time. It's more like a whiteboard that gets erased at the end of every day. The only way anything persists is if someone writes it down somewhere before the erasing happens. Your configuration files (CLAUDE.md), your knowledge base, your handoff documents — those are the writing-it-down. Without them, every session starts from zero.

The Three Types of Forgetting

Click any card to see what it is, why it happens, and what you can do about it.

🔴
Type 1: Session Boundary
The hard reset. Everything gone.
Total loss
🟡
Type 2: Compaction
Emergency summary. Partial loss.
Partial loss
🔵
Type 3: Attention Fade
Nothing lost. Just ignored.
Quality degradation

Type 1: Session Boundary — The Hard Reset

What happens: When a session ends, the entire context window is wiped. Claude starts the next session with zero memory of what happened before. It literally does not know you've ever talked before.

When it happens: Every time a session ends — whether you close the app, the session times out, or you start a new conversation.

What's lost: Everything that only existed in the conversation. Decisions made, context established, preferences stated mid-session, work in progress. All of it.

What survives: Only what's been written to persistent storage before the session ends. Your CLAUDE.md loads automatically next session. Files you created or edited persist. Anything you explicitly saved to a knowledge base persists. Anything else? Gone.

The fix: Write things down before the session ends. Session logs, decision records, updated task lists, key context files. If it's not in a file, it doesn't exist next session. Period.

Type 2: Compaction — The Emergency Summary

What happens: When the context window fills up during an active session, the system automatically summarizes the older conversation to free space. This is called compaction. The summary replaces the original messages.

When it happens: Automatically, without warning, when the context window approaches capacity. You'll sometimes see a note like "This conversation was summarized" — but not always.

What's lost: Specific details, exact wording of instructions, nuanced context, specific numbers and dates, anything that wasn't prominent enough to make it into the summary. The summary is written by a model — it makes editorial choices about what matters.

What survives: High-level decisions, major tasks, the general direction of work. The "what" survives better than the "how" or "why."

Why it's sneaky: Claude doesn't tell you what was lost. It continues working from the summary as if nothing happened. You only notice when Claude gets something wrong that you told it earlier.

The fix: Write critical state to files before the window fills up. Don't wait for compaction — by then it's too late. Good rule of thumb: if you've been working for 2+ hours, write down your current state even if nothing seems wrong.

Type 3: Attention Fade — The Gradual Drift

What happens: Nothing is actually deleted. Your CLAUDE.md instructions, rules files, and earlier messages are still in the context window. But as the conversation grows, they're buried under thousands of tokens of newer content — and Claude's attention drifts toward the most recent content.

When it happens: Gradually, over the course of a long session. You won't notice a specific moment — it's a slow drift. By message 40, Claude may be less reliably following instructions that it followed perfectly at message 5.

What's affected: Formatting rules (suddenly using em dashes again), behavioral guidelines (getting more verbose), role-specific constraints (doing things it was told not to do). The rules are still there — they just have less weight than they did at the start.

What's NOT affected: The actual work content, facts, what you've built together. This type of forgetting is about behavior, not knowledge.

The fix: Restate the key rule in your message. "Remember, no em dashes" in your current message hits harder than line 47 of a rules file. Or start a fresh session — a new session gives every rule full attention weight again.

See Compaction in Action

Here's what happens to a real conversation when compaction kicks in. The left panel is what actually happened. The right is what Claude sees after compaction.

Before Compaction — What Actually Happened

You
Always use Celsius for temperatures. Important — I'm in Canada.
You
Also, never suggest products over $50 CAD.
Claude
Got it — Celsius and under $50 CAD. I'll keep those in mind throughout our conversation.
You
Great. Let's start planning the camping trip. We're going to Algonquin Park in Ontario.
Claude
Algonquin in late summer averages 22°C during the day, dropping to 10°C at night...
You
[30 more exchanges about camping gear, meal planning, routes...]
You
What sleeping bag do you recommend?

After Compaction — What Claude Now Sees

AI-Generated Summary (replaces all earlier messages)
The user is planning a camping trip to Algonquin Park in Ontario. Discussion covered gear options, meal planning, and route selection. The user asked about sleeping bag recommendations.
Gone: "Always use Celsius." Gone: "$50 CAD limit." Gone: The exact gear discussed. Gone: Specific route decisions.

Survived: High-level topic (camping, Algonquin). General context (Ontario). Current question (sleeping bag).
Claude now recommends a sleeping bag rated to 15°F (-9°C)... for a price of $79.99 USD... [Both constraints violated — not malice, just lost context]
The cruel irony: The most important instructions are often given early in a conversation ("always do X," "never do Y"). These are exactly the messages most likely to be summarized away during compaction. The summary captures what happened — not what you asked Claude to remember.

Attention Fade — Watch It Happen

Drag the slider to simulate conversation length. Watch how much "attention weight" Claude gives to different parts of the context.

Conversation length: 10 exchanges Early session
Your current message
95%
Last 3 exchanges
80%
Rules files (hard rules)
75%
CLAUDE.md identity
72%
System prompt
70%
Instructions from hour 1
65%
First message you sent
60%

What Survives What

Type of Content Session Boundary Compaction Attention Fade
Your CLAUDE.md rules Survives (reloads) Survives (always present) Weakens over time
Decisions made in conversation Lost Partially summarized Still present
Files you created or edited Survives (on disk) Survives (on disk) Survives
Specific instructions ("use Celsius") Lost Often lost in summary Weakens with distance
Knowledge base / search docs Survives (on server) Survives (not in context) Survives
Work in progress (unwritten) Lost High-level summary only Still present
Exact numbers, dates, specifics Lost Usually lost in summary Weakens with distance
Formatting preferences Survives (in CLAUDE.md) Survives (in CLAUDE.md) Drifts in long sessions

The Survival Guide

One fix per type. Match the problem to the solution.

For Session Boundary

  • Write key decisions to files before ending
  • Keep a running session log
  • Put persistent rules in CLAUDE.md, not just chat
  • Use a handoff document between sessions
  • Save critical state to a knowledge base

For Compaction

  • Write critical state every 1-2 hours proactively
  • Don't rely on early-session constraints for late work
  • Restate constraints in later messages if they matter
  • Keep important rules in CLAUDE.md (survives compaction)
  • Start a fresh session for new major topics

For Attention Fade

  • Restate the rule in your current message
  • Start a fresh session to reset attention weights
  • Keep the most critical rules at the TOP of your rules files
  • Use clear, emphatic language for critical rules
  • Notice drift early and correct it immediately

The Core Principle

If it's not written somewhere Claude can find next session, it doesn't exist.
This isn't pessimism about AI. It's the architecture. Claude is not building a relationship with you that persists over time. It's loading a briefcase of context at the start of each session. Fill the briefcase intentionally and everything works. Leave it up to chance and you'll keep explaining the same things over and over.
Why power users obsess over their config files: They're not being nerdy. They're building the briefcase. Every line in CLAUDE.md, every document in the knowledge base, every structured handoff — that's the institutional memory that survives the nightly reset. The session ends. The files don't.

Next: Phase 1, Session 5

How Claude "Reads" Instructions — Attention, priority, and why where you put something matters as much as what it says.

We'll look at the difference between instructions at the top vs. the bottom, in CLAUDE.md vs. in your message, and why emphasis markers like CAPS and bold do less than most people think.