"I just corrected that — why is it doing the same thing again?"
If you've spent any time with Claude, you've probably felt this frustration. You ask it to stop using bullet points, it agrees, and then three messages later the bullet points are back. You point out a factual error, it apologizes and fixes it — then makes the exact same mistake in a different part of its response. You spend more time correcting Claude than actually getting work done.
This isn't Claude being stubborn or ignoring you. It comes from a fundamental characteristic of how large language models work — and once you understand it, the fixes become obvious. This guide explains what's actually happening, and then gives you concrete techniques to make your corrections actually stick.
Why Claude "Forgets" What You Just Told It
The short answer: Claude has no long-term memory between conversations, and even within a long conversation, early instructions gradually lose their influence.
Claude works by reading everything visible in the current conversation window — your messages, its responses, any files you've shared — and generating a reply based on all of that context. This is called the "context window." It contains everything Claude knows about your current session.
Two situations cause corrections to fail:
Situation 1: The conversation gets long
As a conversation stretches to 50, 100, or more exchanges, your early instructions end up buried deep in the context. When the context window approaches its limit, older content gets compressed or summarized automatically. The nuanced correction you made in the first ten messages may survive only as a vague summary — not the specific rule you were trying to set.
This is different from human memory. When a person forgets something you told them, it usually means they weren't paying attention. When Claude "forgets" an instruction, it means the instruction has moved beyond the range of text that Claude is actively weighing in its response generation. The mechanism is architectural, not attentional.
Situation 2: You start a new chat
The moment you open a new conversation, Claude has zero memory of what happened before. Everything you patiently explained in your last session is gone. This is by design — it protects privacy and prevents one session's context from contaminating another.
Many users assume that because Claude is sophisticated, it must be learning their preferences over time. It isn't — at least not within the standard Claude.ai interface. Each conversation starts completely fresh.
Temporary Corrections vs. Permanent Fixes
Once you understand this, it becomes clear that there are two fundamentally different kinds of correction:
Temporary corrections (valid for this conversation only):
- "Please don't use bullet points"
- "Make the tone more casual"
- "That calculation was wrong — here's the correct approach"
These work fine within a single session, but they fade as the conversation grows longer, and they vanish completely when you start a new chat.
Permanent fixes (valid across all sessions):
- Writing rules into Claude's profile settings (custom instructions)
- Writing rules into a Project's system prompt
- Starting each conversation with a dedicated instruction block
The right approach depends on how important and how reusable the correction is. A one-off style preference for today's document belongs in the conversation. A rule about how Claude should always respond to you belongs in your profile settings.
Three Immediate Techniques That Work Today
Technique 1: Front-load your rules at the start of each conversation
Instead of correcting Claude reactively mid-conversation, move your instructions to the very first message before the actual task.
# Rules for this session
- Write in flowing prose — no bullet points or numbered lists
- Keep formal register throughout
- Flag any assumptions you're making explicitly
- Ask clarifying questions before starting long tasks
With these rules in mind, please help me with the following:
[actual request here]
Creating a personal "rule block" you paste at the start of relevant conversations is one of the highest-leverage habits you can build. Instructions at the top of a conversation carry more weight than corrections made in the middle of one — they're part of the initial framing, not an afterthought.
I keep a few of these blocks saved in my notes app and paste the appropriate one depending on what I'm working on. It takes ten seconds and eliminates most correction cycles entirely.
Technique 2: Always explain why when you correct
"Don't do that" is easy to forget. "Don't do that because of X" is much harder to lose track of. When Claude understands the reason behind a rule, it can apply that reasoning to similar situations it hasn't explicitly been told about.
✗ "Stop using bullet points."
✓ "Please avoid bullet points in this project. The content will be pasted
directly into a Word document where Markdown doesn't render, so bullet
symbols appear as literal asterisks. Write lists as flowing sentences instead."
The explanation helps Claude generalize the correction — it will now apply the same logic to numbered lists, bold text, and other Markdown formatting it might otherwise reach for. Without the reason, Claude knows one specific thing to avoid. With the reason, it understands a whole category of things to avoid.
Technique 3: Pause long conversations to summarize the rules
When a conversation has gone on for a while and you notice corrections starting to slip, try this reset prompt:
Before we continue, please list the key rules and constraints we've
established in this conversation. Then confirm you'll apply them
consistently from here on.
Having Claude articulate the rules itself pulls them back into active focus. This works particularly well during long editing or writing sessions. I find that doing this every 30 to 40 exchanges prevents the gradual drift where instructions slowly erode as the conversation grows.
The Permanent Solution: Claude Projects
For rules you want to enforce across every conversation — not just the current one — Claude Projects with custom instructions is the most powerful option available.
Projects let you write a system prompt that automatically applies to every conversation within that project. You set it once; it runs every time without any extra effort.
To set this up:
- Open Claude.ai and select "New project" from the left sidebar
- Name the project and add a brief description of its purpose
- Open "Project instructions" and write your standing rules
For a writing project, this might look like:
# Writing Standards
## Style
- Write in flowing paragraphs, not bullet points or numbered lists
- Keep sentences under 25 words where possible
- Use active voice by default
## Content requirements
- Include at least one concrete example per major point
- Define technical terms on first use
- Never end responses with "I hope this helps" or similar phrases
## What to avoid
- Filler phrases: "It's worth noting that...", "Importantly...", "It's crucial to..."
- Summaries that just repeat the introduction
- Hedging every sentence with "may", "might", or "could potentially"
Once this is in place, you never have to repeat these corrections. Claude applies them automatically, even in conversations you start weeks later. You can create multiple Projects for different use cases — one for writing, one for coding work, one for research — each with its own tailored rules.
Your Profile Settings: The Most Persistent Option
Beyond Projects, Claude.ai has a profile-level custom instructions feature that applies across all conversations, including those outside of any Project.
To access it:
- Click your profile icon in the top right corner of Claude.ai
- Open "Profile & Settings"
- Find the "Custom instructions" field
This is the right place for your most universal preferences — the things you want from Claude in every possible context.
- Always respond in English unless I write to you in another language
- Don't end responses with "Is there anything else I can help you with?"
- When you're uncertain, say so explicitly rather than presenting guesses as facts
- Keep explanations concise — I prefer clarity over completeness
If there's a correction you find yourself making in every single session, that's a strong signal it belongs in your profile settings rather than in each individual conversation.
When a Specific Conversation Feels Especially Forgetful
There's one pattern that causes corrections to fade faster than usual: topic shifts within a long conversation.
Say you're drafting a formal email in English, then ask Claude to translate a version into French, then go back to working on the English email. The tone and style rules you established at the start — written with formal English in mind — can quietly get dropped during the language shift. Claude may not know which rules were specific to English and which were meant to apply throughout.
When you change topics or modes mid-conversation, explicitly state which rules carry over:
Now let's work on the French translation. The "no bullet points" rule still
applies here, but the formal English register instruction obviously doesn't —
use standard professional French style instead.
This kind of explicit handoff prevents Claude from silently resetting your preferences when the conversation changes direction. It takes one extra sentence and saves you several rounds of correction.
A Pattern Worth Noticing
After using Claude heavily for a while, I've noticed something useful: the corrections that keep coming back are the ones I care about most. That's actually valuable data. If you're correcting the same thing three sessions in a row, that preference belongs in your profile settings or your Project instructions — not in the conversation.
The workflow I've settled on is roughly this:
First time correcting something: do it in the conversation with a clear reason. Second time correcting the same thing: add it to the relevant Project's instructions. Third time: add it to my profile settings.
Following this pattern, my profile settings and Project instructions have gradually become a well-tuned reflection of how I actually want Claude to work. The correction cycles have mostly stopped — not because Claude got better at remembering, but because I moved the important rules somewhere Claude will always see them.
Summing Up: Corrections Need a Permanent Home
The core insight is this: "Claude keeps making the same mistake" is almost never an ability problem. It's a placement problem. Your correction is valid — it just isn't stored anywhere that Claude can reliably access across time.
Think of it in three tiers:
- Rules stated mid-conversation → active now, fades with length
- Rules at the start of each conversation → stable for the session
- Rules in Projects or profile settings → persistent across all sessions
The more important a rule is to your workflow, the higher tier it belongs in. Start by moving your most repeated correction into a Project's custom instructions. That single change is usually enough to break the cycle.