Engram vs ChatGPT and Claude built-in memory

What is the difference?

ChatGPT and Claude both ship built-in memory now. They watch your conversations, extract facts about you, and reuse them later. It costs nothing, requires no setup, and for a lot of people it is enough.

Engram is a different shape. Your memory is the notes you already write, stored as markdown files you own, and any AI reads them over an open protocol. There is no extraction step between what you wrote and what the AI reads.

What they do well, honestly

Their memory is genuinely good and it has gotten better fast.

• Zero setup. It is on by default and it works.
• Both ship viewers now. Claude replaced its rolling summary with individually editable Topics in July 2026, with export and import. ChatGPT has an editable saved-memory list, a revisable profile summary, and shows per-answer memory sources so you can see which memory influenced a reply.
• Enormous distribution. It is already in the app you already opened.

If you use exactly one assistant, do not write notes, and want no moving parts, their memory is the right tool and you should use it.

Side by side

	Assistant built-in memory	Engram
Cost	Included	Free tier, then $7 or $14 per month
Setup	None	Connect a vault, connect a client
How memories get written	An LLM extracts facts from your chats	You write notes. Your AI can also write notes.
Can you read the stored form	Yes, as a list of extracted facts	Yes, it is a markdown file
Can you check it against a source	No, the source was a conversation	Yes, the note is the source
Can you fix a wrong one	Edit the extracted text	Open the file and edit it, in any editor
Portable to another AI	No	Yes, plain files over an open protocol
Shareable with another person	No	Yes, it is a folder
Works offline in another tool	No	Yes, they are files on your disk

The real difference is verifiability, not visibility

It would be easy and wrong to say their memory is a black box you cannot inspect. That was true eighteen months ago. It is not true now: both ship viewers, both let you edit and delete, and ChatGPT will even tell you which memory it used for a given answer.

The difference that survives is one step earlier. In those systems an LLM reads your conversation and writes a paraphrase of what it thinks is true about you. You can read the paraphrase. You cannot check it, because the thing it paraphrased was a conversation that no longer exists in a form you can diff against.

That matters when the extraction is wrong, which is a documented failure mode across this category. Irony and sarcasm get recorded as sincere. Details get merged. A correction does not always displace the thing it was correcting. A record of when a wrong fact was written is not the same as being able to see that it is wrong.

Engram has no extraction step. The memory is the note. If it says something wrong, you open the file and fix the line, in Obsidian or vim or anything else, and that corrected file is what the AI reads next time. We call this provenance rather than transparency, because the point is not that you can look at it. The point is that there is a source to check it against.

The portability difference

Assistant memory is per-user and per-vendor by construction.

Claude’s memory is tied to your account and is account-wide rather than per-project, so work context and personal context share one bucket. On Team and Enterprise plans an owner can switch memory on or off for the organization, but individual memories stay individual. ChatGPT’s works the same way: yours, in their app.

Neither can be handed to a colleague. Neither moves when you switch assistants. The best model changes every few months, and each switch starts you over.

Engram’s memory is a folder. It moves because folders move.

What Engram does not do

We are not going to pretend this is a clean sweep.

• Engram does not watch your conversations. If you want memory to accrue without you doing anything, that is exactly what their memory does and ours does not. Ours holds what you write, plus what you let your AI write back.
• It costs money above the free tier. Theirs is included.
• It is more setup. A vault, a connection, a client.
• The hosted service is not end-to-end encrypted, because semantic search requires the server to read note text. Self-host if that matters to you.

Can I use both?

Yes, and most people should. Let the assistant remember your tone preferences and small talk. Keep the things you would be annoyed to lose in files you own, and point every assistant at those.

Summary

Their memory is free, automatic, and good at what it does, and it belongs to them. Yours should be the notes you already write: checkable against a source, fixable by hand, readable by whichever AI you use next year, and shareable with a person. Their memory keeps you. Yours follows you.