In the last lesson you learned that every AI session carries a backpack full of tokens. Now the uncomfortable question: what happens when the backpack gets too full? The answer explains one of the most confusing experiences beginners have with AI — the session that starts brilliantly and slowly turns vague, forgetful, and stubborn. It is not your imagination. Long conversations genuinely get worse.

The window is not the memory

Two terms get mixed up constantly. Keep them separate:

A useful comparison: the context window is the desk; memory is the filing cabinet. You can only work with what fits on the desk. A bigger desk helps, but a desk buried under yesterday's coffee cups, dead drafts, and three contradictory to-do lists is not a better desk. It is a messier one.

That is why a larger context window, by itself, does not fix sloppy context habits. It just gives you a larger surface to clutter.

The five signs of a crowded session

Watch for these. Any one of them means the backpack from Lesson 9.1 is too full:

1. Crowding. The useful signal is buried under exploration, pasted logs, and dead ends. The model starts quoting irrelevant old material or missing the file right in front of it. 2. Stale instructions. An early direction ("use the old pricing page") still sits in history after you reversed it ("actually, rebuild it"). The model keeps obeying the ghost. 3. Conflicting decisions. You decided A, then B, then "actually A again." All three versions are in the chat. The model blends them into something you never chose. 4. Lost priorities. The original goal ("fix mobile overflow") drowns under side quests ("also, while you're here…"). The model optimizes for the latest chatter, not the real task. 5. Long-session drift. After dozens of turns, answers get vaguer, more agreeable, and less anchored to your files. The session is running on fumes and momentum.

Bloated chat: 80 messages, 4 reversals, 2 giant logsCrowding + stale instructions + driftSlower, vaguer, more expensive answersCompact handoff + clean sessionFocused task + fresh attentionFaster, sharper, cheaper answers

The left path is the default. The right path is a skill, and you already met its raw materials in Class 02: README.md, CONTEXT.md, SPEC.md, and handoff files.

The reset loop: save the decisions, drop the noise

When a session reaches a natural stopping point — a feature works, a decision is made, the chat passes ~20–30 turns, or you spot any symptom above — run the reset loop:

Work → checkpoint → save durable decisions in Markdown
→ open clean session → read context files → continue

Concretely:

1. Work until a checkpoint: something builds, a decision settles, or focus starts slipping. 2. Checkpoint out loud: "What is done, what did we decide, what is next?" 3. Save the durable answers in a short HANDOFF.md (template below). Carry decisions and evidence — not every dead end. 4. Open a clean session. Close or abandon the bloated one. 5. Read context files first: CONTEXT.md, SPEC.md, and the latest handoff. Then give the new session *one* task.

Put handoffs in one predictable place — docs/handoffs/, notes/handoffs/, or project-memory/ — so the next agent can find them without asking where the last session left off.

Exercise: turn a messy conversation into a HANDOFF.md

Imagine (or open) a real project chat with at least 15 messages, one reversed decision, and one pasted log or error. Distill it into this file: docs/handoffs/<topic>.md.

# Handoff: <short topic>

## Goal
[One sentence: what is this piece of work trying to achieve?]

## Completed
- [What works now, with evidence: build passes, test name, file changed]

## Decisions made
- [Decision + why. Drop the debate, keep the conclusion.]

## Current state
- Branch: [name]
- Changed files: [list]
- Known issue: [what is still broken or unreviewed]

## Next task (one task only)
[The single next action, with a stop point.]

## Do not do
- [Boundary 1: e.g. do not change the schema]
- [Boundary 2: e.g. do not deploy to production]

Then start a genuinely new session with this opener:

Done looks like: one handoff file under 40 lines that a stranger could act on.

Quick check: hand it to the test — could a new agent complete the next task without reading the old chat? If it still needs the old chat, the handoff is missing a decision or a file path.

Common mistake: copying the whole conversation into the handoff "just in case." That re-inflates the backpack you just emptied. Summarize conclusions; link to long evidence instead of pasting it.

Check your understanding

1. What is the difference between a context window and long-term memory? 2. Name two symptoms of context crowding and what causes each. 3. Why does a bigger context window not automatically fix a drifting session? 4. At what point should you run the reset loop instead of continuing the chat? 5. What belongs in a handoff file, and what should deliberately be left out?