Your Claude Sonnet 5 memory is intact: the saved facts and preferences survived the default swap, because on consumer Claude they live at the account level, not inside any single model. The unsettling part is quieter. You spent weeks tuning Claude, and a model you never chose is now the one reading those notes, so tone, verbosity, and judgment can shift even when the stored data is byte-for-byte identical.
Your context is a tenant in someone else's building, and the landlord just renovated without telling you. The data stayed. The mind reading it is a stranger now.
Anthropic made Claude Sonnet 5 the default model for Free and Pro plans on June 30, 2026, with access also open to Max, Team, and Enterprise users. Anthropic moved most free and paid consumer users onto it without an opt-in. So the assistant you talked to last week is not exactly the assistant you are talking to now, and nobody asked you first.
What Claude Sonnet 5 actually is
Claude Sonnet 5 is Anthropic's newest mid-tier model, pitched as the most agentic Sonnet yet: it can make plans, use tools like browsers and terminals, and run autonomously at a level that a few months ago needed larger, pricier models. That is the claim in nearly every launch write-up, and the benchmarks back the direction of travel. On SWE-bench Pro, the harder agentic-coding variant, Sonnet 5 scores 63.2 percent versus Sonnet 4.6's 58.1 percent.
On the API, Anthropic set introductory pricing of 2 dollars per million input tokens and 10 dollars per million output tokens through August 31, 2026. After that, standard rates kick in at 3 dollars in and 15 dollars out per million tokens. The framing is deliberate: a cheaper way to run agents, which is why coverage focused on cost per task rather than raw intelligence. For most readers, though, that pricing is trivia. The consequential fact is the default swap.
Free and Pro users who open Claude and start typing are now getting Sonnet 5's judgment, its refusal patterns, its verbosity, and its interpretation of your instructions, whether or not they know a model changed underneath them. Nobody sent them a changelog to acknowledge first.
What happens to your Claude Sonnet 5 memory
Nothing happens to the memory itself, and that is the point people miss. Claude's consumer memory is a generated summary, not a raw transcript. Claude automatically summarizes your conversations and builds a synthesis of key insights across your chat history: your role, your recurring topics, your stated preferences. That synthesis refreshes about every 24 hours and gets injected as context into every new standalone conversation. If you want the fuller mechanics, we cover them in our guide to how Claude memory works.
The swap-specific detail is where that synthesis is stored. It lives at the account level, and each project keeps its own separate memory space and dedicated summary. Anthropic rolled memory out to every tier, including free accounts, in early March 2026. Because the synthesis sits against your account rather than baked into model weights, replacing the default model does not touch it. The facts are still there on July 1 exactly as they were on June 29.
What carries over, and what gets re-read
Two things happen when a new default model reads your memory. First, the stored synthesis carries over untouched: same facts, same preferences, same project summaries. Second, a different model re-interprets that synthesis from scratch on every new chat. Memory is data. Behavior is the model applying that data. Swap the model and you keep the data but change the application.
There is a subtler layer for anyone using Claude on long documents or agent runs. When a conversation approaches the context window limit, Anthropic's compaction can summarize older context server-side so the session keeps going. That summarization is model-driven too. A new default model decides what to keep and what to drop differently, so mid-session memory can compress along slightly different lines than it used to.
If Claude suddenly forgets a detail mid-chat that it clearly knew earlier, that is usually compaction dropping it during summarization, not your account memory failing. Restate the detail, or move it into Profile Instructions where it is always loaded verbatim.
The wrong axis: what most coverage missed
Nearly every Sonnet 5 write-up argued about whether it is smarter, faster, or cheaper than Sonnet 4.6. That is the wrong axis. Here is the sharper version you can quote to win the argument: your memory and your working context live inside one vendor's product, and that vendor can change the model reading them at any time, silently, without your consent and without a changelog you have to sign off on.
Consider what a default swap does to a workflow you tuned. You taught Claude your preferences through weeks of corrections. You wrote Profile Instructions. You built up a memory summary that finally made it answer the way you wanted. That is real work, and it is exactly the kind of thing loss aversion makes you dread losing. None of the data disappeared. But a new model interprets it, so the same instruction that produced a tight three-line answer last month might now produce a cautious eight-line one, or the reverse. You changed nothing. The floor moved.
This is not an Anthropic-specific complaint. OpenAI and Google do the same thing when they rotate defaults, and a similar dynamic shows up when an embedding model update quietly breaks AI memory. The pattern is structural: when your assistant's memory is a vendor feature, the vendor owns the read path. You own the words you typed. They own how those words get turned into behavior, and they can change that overnight.
Why a silent model swap changes the assistant's behavior
A more agentic default model changes defaults you never set. Sonnet 5 is tuned to plan, to use tools, to check its own output without being asked, and to finish multi-step tasks where earlier Sonnet models stopped short. Those are real improvements for agent workloads. For a casual chat, they can read as the assistant being more verbose, more proactive, or more willing to take independent steps than you wanted.
- Tone and length shift. A new default can run more or less terse against the same Profile Instructions, so answers feel off even though your settings did not change.
- Refusal boundaries move. Safety tuning differs between models. A request Sonnet 4.6 answered might now get a caveat or a decline, and the reverse.
- Instruction-following changes. How strictly the model honors saved preferences versus the immediate prompt can shift, changing which one wins when they conflict.
- Autonomy rises. A more agentic model may take extra steps or call tools you did not expect, because acting independently is closer to its default.
- Memory compression differs. Server-side summarization of long chats can keep and drop different details, so what the model recalls mid-session subtly changes.
None of these are bugs. They are the honest consequence of one component being replaced under a system you depend on. The assistant did not forget you in the sense of losing data. It simply got a new brain reading the same notes, and the landlord did not knock first. The fix is not to fight the new model. It is to notice which parts of your setup are yours to keep and which parts the vendor can rewrite on its own schedule, then move the important parts to the side you control.
Model-locked memory vs a memory layer you own
| Property | Claude Sonnet 5 memory (vendor feature) | External memory layer (yours) |
|---|---|---|
| Where your data lives | Inside Anthropic's product, tied to your Claude account | In a store you control, outside any single assistant |
| Survives the Sonnet 5 default swap | Data survives; behavior reading it can change | Data and its structure stay identical regardless of model |
| Works across assistants | Claude only; no reach into ChatGPT or Gemini | Model-agnostic; the same memory feeds whichever assistant you ask |
| Who decides how it is summarized | The vendor's current default model, on its schedule | You keep the source; you decide what gets condensed |
| If you leave the vendor | Memory synthesis does not port out with you | Your memory stays because it was never the vendor's to hold |
The comparison is not anti-Claude. Claude Sonnet 5 is a strong model and, for many people, a genuine upgrade. The point is narrower: relying on a single vendor's memory means your continuity is only ever as stable as that vendor's current default, and defaults change on the vendor's timeline, not yours. The more of your working life you route through one assistant, the more a routine model rotation quietly reshapes your daily output. This is also why Claude memory versus projects can feel like managing many brains rather than one.
The MemX angle: keep the memory, swap the model
This is the exact problem MemX is built for. MemX is a model-agnostic external memory layer: your documents, chats, voice notes, and photos live in one place you control, and you ask questions across them in plain language. It sits above whichever assistant you use, so when Anthropic makes Sonnet 5 the default, or OpenAI or Google rotate theirs, your memory does not move and does not get re-summarized by someone else's new model. To borrow the earlier framing, you own the building instead of renting it, so no landlord can renovate your context overnight.
On privacy, MemX is private by architecture: per-user isolation, customer-managed keys, encryption at rest, and on-device processing. That is deliberately not a claim of end-to-end encryption or zero knowledge, because MemX is neither, and being straight about that matters more than the marketing. What it does give you is a memory that answers to you instead of to whichever model happens to be the default this month. If a silent Sonnet 5 swap made your assistant feel like a stranger, owning the memory layer is the part of the stack you can actually keep steady.
What to do this week
- Check which model you are on. Free and Pro now default to Sonnet 5, so if Claude feels different, the model likely changed, not your memory.
- Re-read your Profile Instructions. Preferences that were implicit under the old default may need to be stated explicitly for the new one to honor them.
- Move must-keep facts out of chat memory. Instructions loaded verbatim beat a 24-hour synthesis that a new model re-interprets.
- Watch long chats for dropped details. Compaction summarizes differently under a new model; restate anything critical if it goes missing mid-session.
- Keep a copy of what matters outside the vendor, so the next default swap does not decide your continuity for you.
01Does Claude Sonnet 5 have memory, and did it change mine?
Yes, Claude Sonnet 5 uses the same account-level memory. Making it the default on June 30, 2026 did not change your stored facts or preferences. Only the model reading that memory changed, which can shift how answers feel.
02Did Claude Sonnet 5 delete my memory?
No. Consumer Claude memory is stored at the account level, not inside the model, so the Sonnet 5 default swap left your saved facts and preferences intact. The model reading that memory is the only thing that changed.
03Why does Claude feel different after the update?
Claude Sonnet 5 is a new default model, so it interprets your saved memory and instructions differently. Tone, length, refusal boundaries, and autonomy can shift even though your stored data and settings did not change.
04Is Claude Sonnet 5 the default for free users?
Yes. Anthropic made Claude Sonnet 5 the default for both Free and Pro plans on June 30, 2026. It is also available to Max, Team, and Enterprise users across those higher tiers.
05Can I move my Claude Sonnet 5 memory to ChatGPT or Gemini?
Not directly. Claude's memory synthesis is a vendor feature and does not port to other assistants. To keep memory that works across models, you need an external, model-agnostic memory layer that you control.
About the author: Aditya Kumar Jha is a founding software engineer at MemX, where he works on model-agnostic memory and retrieval. He tracks model releases and memory behavior across ChatGPT, Claude, and Gemini, and writes about what actually changes for users when vendors rotate their defaults. Before publication, he verified every date, version, and price here against Anthropic's launch announcement and primary reporting.
