Context Rot

A team spends a whole day in a planning workshop. By five o'clock the whiteboard holds three versions of the timeline, a budget everyone rejected at lunch, and arrows to ideas that were crossed out but not erased. Someone walking in could not tell which plan is current. Even the people who were there start quoting the rejected budget.

Long sessions with an AI assistant go the same way. The first hour is sharp. Two hours and a hundred messages later, it forgets a rule you set early on, brings back a fix you already rejected, and mixes the old version of your code with the new one. The context has rotted.

#Why long sessions degrade

Every message, every attempt and every tool result stays in the context and is reread on each call. As the session grows, several things happen at once.

• Dead ends stay. A wrong approach you abandoned is still there, in full, looking just as authoritative as the right one. The model cannot tell which one you gave up on unless that is clearly said.
• Old versions pile up. Three drafts of a function or an email sit in the context. Asked to change "the function", the model may edit the wrong one or blend them.
• Early instructions fade. A rule from message 3 is now a tiny share of a huge context, buried under everything said since. As The Context Window showed, models tend to use a long context less reliably than a short one, and near the limit the app may trim the rule away entirely.
• Mistakes become examples. If the model made the same error twice earlier, those replies are now part of the text it continues. It tends to keep the pattern.

#Watch it happen

A developer fixes a bug with an assistant. The token counts are made up. Track how much of the context is still current.

each turn: context += new message + replynothing is ever removedfresh session: context = handoff + current code

The handoff kept one line from the dead end: that it failed, and why. That line stops the fresh session from trying it again. The 800 tokens of A's code are gone.

#Signs your session has rotted

• It forgets or breaks a rule you set early on.
• It brings back an approach or a bug you already rejected.
• It edits an old version of something instead of the latest.
• You are writing more corrections than requests.
• Replies get longer, vaguer, and more hedged than at the start.

Two or more together mean it is time to change the context, not argue.

#What to do about it

1. Start fresh with a handoff. Ask the assistant for a short summary, correct it yourself, and paste it into a new chat. A good handoff has the goal, the current state, the decisions and rules, what was tried and failed with the reason, and the next step.
2. Keep a notes file. For work that spans days, keep the handoff in a file outside the chat: a project brief or a NOTES.md next to your code. Load it at the start of each session and update it at each milestone. Some coding tools load such a file automatically. It is memory that you control.
3. Edit instead of piling on. If a reply went wrong because of your message, many apps let you edit that message and regenerate. The bad reply leaves the context instead of staying as an example.
4. Split the work. One task per session. A tangent about logging deserves its own chat.
5. Restate what matters. When a late request depends on an early rule, repeat the rule in the request. It costs one sentence.

# Date bug fix: notes
 
Goal: dates entered as DD/MM must not be read as MM/DD.
Rules: keep the public API unchanged. All tests must pass.
 
Current state: parser fixed in parseDate() (attempt B). Tests pass.
Tried and failed:
- Rewriting the parser (attempt A): changed the public API.
 
Next: add tests for 29 Feb and for two-digit years.

Some apps do part of this for you, often called compaction: when a session nears its limit, older turns are replaced by a summary, and old tool results may be cleared. It helps, but the summary decides for you what matters, so read it for important work.

#Common mistakes

#Try it yourself

#Check yourself

#What to practise next

Next time a session starts to drift, write the handoff and start fresh, and compare the next reply. Then move on to What Is an Agent, where a model runs long tool loops on its own and managing context becomes the core of the design.