What this post covers. ① Turning Claude memory on (30 seconds). ② Five things to pin explicitly in the first 5 minutes. ③ Editing memory mid-conversation. ④ Global vs project memory. ⑤ Importing memory from ChatGPT and Gemini. ⑥ Debugging memory — search, pause, reset.
“Claude doesn’t seem to remember how I work.” I hear this a lot. People end up retyping the same context every session, get fed up, and turn the memory feature off entirely. Honestly, I did the same thing at first. The thing is, memory isn’t a “toggle it on and it learns” feature. There’s an automatic Memory Synthesis pass that runs roughly every 24 hours, but waiting on that alone is slow. Spend 5 minutes pinning things explicitly and they take effect immediately.
This post is a guide to those first 5 minutes. As of April, with the feature live on the free plan, this is the fastest way to get memory actually working for you.
First, turning it on (30 seconds)
Top-right profile in Claude → Settings → Capabilities → Memory toggle ON. That’s it. Available on every tier from Free up through Pro, Team, and Enterprise (as of March 2, 2026).
One thing to watch out for: conversations in Incognito Chat do not contribute to memory. So if you have a sensitive conversation you don’t want piling into memory, run it in incognito and you’re safe.
Five things to pin explicitly in the first 5 minutes
Once memory is on, open a new chat and tell Claude these five things in one go. Start with “Remember this:” and Claude updates the memory summary right then. No 24-hour wait — it’s live in your next conversation.
① Role and domain
Something like “I’m a backend developer. I mostly work in Python, Node.js, and PostgreSQL.” Claude calibrates response length, example code language, and assumed knowledge level differently based on this.
② Response tone
“Lead with a conclusion under 200 characters, then add detail below if needed.” Or rejection-style options like “Don’t use emoji.” That way you stop nagging Claude about the same thing every conversation.
③ Languages and tool stack you use
“For code examples, prefer TypeScript. I use Vite for builds and Vitest for tests.” Without this, Claude will keep defaulting to React + Webpack circa 2019, which gets old fast.
④ One or two projects you’re working on
Something like “I’m building an auto-publishing pipeline for my personal blog. Tistory API + Python.” Even one line gets you automatic continuity — next conversation, Claude will frame answers around “your blog automation project.”
⑤ Rejection options
“Don’t both-sides every answer.” “Keep explanation minimal outside the code itself.” “Reply only in English.” Spelling out what NOT to do turns out to be surprisingly effective in memory.
Why this works. Auto-synthesis takes time because Claude has to find “patterns worth inferring.” When you state something explicitly, it skips the inference step and lands directly in the memory summary. That’s how you collapse the 24-hour window to seconds.
Editing memory mid-conversation
You don’t need to open Settings. From any chat, drop a line like these and memory updates on the spot.
“Remember this: I prefer Opus over Sonnet for replies.”
“From now on, always include type hints in code.”
“Drop that from memory.” (removal works too)
When the memory summary gets updated, Claude responds with something like “Added to memory.” So you can confirm right in the reply that the update went through.
Conversely, when Claude is using memory in an answer, it cites the source — “Based on the React project you mentioned earlier…” If you don’t see that kind of phrasing, memory isn’t being applied, and you should restate the directive once.
Global memory ≠ project memory
This is one of the questions that keeps coming up on r/ClaudeAI. Claude has two kinds of memory.
| Scope | All chats | Only within that project |
| What to put here | General work style and response tone | Company code context, domain vocabulary |
| Sensitive info | Don’t put it here | Stays isolated inside the project |
The rule of thumb: keep company code context inside projects, and put only general work-style preferences in global. Internal company terminology, API key formats, in-house patterns — if any of that lands in global, it can leak into your unrelated personal work.
On the flip side, generic tone preferences like “don’t use emoji in answers” belong in global. Re-typing them per project gets old fast.
Importing memory from ChatGPT and Gemini
One of the features that shipped alongside the March 2 rollout is importing memory from another AI. Particularly useful if you’ve been on ChatGPT or Gemini for a year or more.
The flow
- In the Claude app, go to
Settings → Memory → Import from another AI - Anthropic gives you a dedicated prompt on screen — copy it and paste into ChatGPT or Gemini
- The other AI dumps its memory as a code block
- Copy the entire code block and paste it into the Claude import box
- Processing takes about 24 hours — work-related context gets prioritized
In my testing, the import handles core stuff — name, role, frequently used languages, response tone — pretty well, but project-level details deserve a once-over after. So even after importing, run the “five things to pin” step above to clean up gaps.
Debugging memory — search, pause, reset
If something feels off about how Claude is responding, head to Settings → Memory and you can edit or delete entries individually. Three options I use most:
- Search (paid plans) — natural-language recall like “continue from that React thread last week.” Lets you reach the original conversation, separate from the memory summary
- Pause — stops accumulating new memory while keeping existing memory active. Good for sensitive work windows
- Reset — permanent delete, no recovery. For when you want a fresh-laptop feeling
Heads up. Reset really means unrecoverable. If you’ve got a rich pile of work context in there, screenshot the Settings → Memory page before you nuke it. Item-by-item deletion is usually enough.
One more thing. Right after I turned memory on, I kept thinking “why doesn’t this feel any smarter?” It took about a week before I noticed auto-synthesis had actually kicked in. That entire week was wasted, and I only realized it after the fact. Five minutes of explicit pinning right after you flip the switch lets you skip that whole week.
The single most impactful thing for me was the rejection options. After pinning “don’t both-sides everything” and “don’t use emoji,” the answer tone shifted completely. Telling AI what NOT to do, explicitly — that’s becoming a real skill in the memory era.
FAQ
Q1. Can I use Claude memory fully on the free plan?
Yes — as of March 2, 2026, the entire memory feature is live on the free tier. Toggling it on, explicit edits, separating global vs project, importing from ChatGPT and Gemini — all available on free. The one exception is past chat search, which is paid-plan only (Pro and up).
Q2. Does what I put in Claude memory get used as training data?
One thing to be aware of. After the 2025 policy change, Consumer plans (Free, Pro, Max) are opted into training-data use by default. Unless you turn it off, conversations and memory can be used for training. Flip Settings → Privacy → Help improve Claude to OFF to opt out, and data is purged within 30 days of the change. Sensitive material — company secrets, customer PII, credit card numbers — shouldn’t go in memory regardless of policy, and to lower the blast radius of any incident, run sensitive work in incognito chat. For reference, Team and Enterprise plans are opted out by default and don’t need any setting changes to stay out of training.
Q3. If I’m switching from ChatGPT to Claude, is the memory import enough?
Not entirely. Import handles the big strokes well — name, role, response tone, frequently used languages — but project-level context often gets dropped or compressed. So right after importing, run the “five things to pin” step above to fill in what’s missing. Auto-synthesis will sharpen things further over the next 24 hours, but having explicit memory in place during that window makes the experience much smoother.
Related links
Feature descriptions in this post reflect Anthropic’s official help docs as of April 28, 2026. UI placement and policy may change later, so double-check the official docs before making decisions.
This post is part of an English mirror of a Korean dev experiments blog. Cross-posted for the global developer audience.