Claude Remembers You. Engram Remembers What You Said.

Claude has memory now, and it's good. It's on by default on Free, Pro, and Max; it saves as you chat; it carries what it knows from the chat app into Claude Cowork; and you can open Settings → Memory and read every single thing it believes about you, topic by topic, and delete any of it. That's a better memory feature than most of this category ships, and it costs nothing.

So the honest question isn't “which one wins.” It's what each one is actually for. Claude's memory is built to remember you — your role, your stack, how you like to be written to. Engram is built to remember what was said — the exact words, in every tool you use, forever.

We build Engram, so treat this the way you'd treat any comparison written by one of the two parties. We've linked Anthropic's own documentation throughout so you can check every claim about Claude against the source.

The 30-second summary

How Claude's memory actually works

Worth being precise, because the mechanism is the whole argument. Per Anthropic's documentation, Claude doesn't store your conversations as memory. It generates memory fromyour conversations — a set of short topic files covering things like your role, your projects, and your communication preferences. Those files live under Settings → Memory, one per topic, and they're deliberately small: Anthropic's own framing is that the files are short, so correcting one pays off everywhere.

That is a genuinely well-chosen design for the job it's doing. Short editable files are auditable, cheap to load into every conversation, and easy to fix when they're wrong. Memory is kept separate per project so your work contexts don't bleed together, and it's shared between the chat app and Cowork.

But notice what the design implies. A topic file is a conclusion Claude drew, not a record of what happened. “Prefers Postgres over MySQL for new services” is what survives; the forty-minute conversation where you worked out why, including the two constraints that would change the answer, is not stored anywhere in memory. Claude's chat search can sometimes find that conversation again — but only if it happened in Claude, and only on paid plans.

Four things Engram does that Claude's memory doesn't

1. It keeps the exact words

Engram stores every message as sent. No summarization pass, no extraction step deciding what was worth keeping, no paraphrase. Search runs over the transcripts semantically and hands back the original text.

This matters in exactly one situation, but it's a situation that recurs: when the thing you need back is the detail nobody thought was important at the time. The precise error string. The version number. The wording of a decision, with its caveats attached. A summary is a bet about what future-you will need, and it's placed by a model that hasn't met future-you. We wrote about why that bet goes bad in Your AI Doesn't Remember Your Conversations. It Remembers a Summary.

2. It crosses vendors, not just apps

Claude's memory travels everywhere Claude goes — web, desktop, mobile, Cowork. That's real portability, and for someone who works only in Claude it may be all the portability they need.

It does not travel to ChatGPT, Cursor, Codex CLI, or anything else, and Anthropic doesn't pretend otherwise. Engram is a Model Context Protocol server, which means it sits on the AI's side of the connection and any MCP-capable client can call it. Research a problem in ChatGPT on your phone, open Claude Code in a terminal that evening, and ask what you found — same memory, one account. That's twelve tools today, and it grows whenever anyone ships MCP support, because there are no per-platform integrations to write.

3. It captures coding sessions automatically

Claude's memory covers the chat surfaces. The Engram plugin for Claude Code captures the work itself — the investigation, the dead ends, the fix that finally landed — verbatim, as you work, retrievable later from any connected AI. For anyone whose real context lives in a terminal rather than a chat window, that's the gap.

4. It has no opinion about what you're allowed to remember

Claude's memory excludes sensitive categories by default — health, religion, politics, identity — and you can opt those back in. A few categories are refused regardless. That's a defensible product decision for a default-on feature shipping to millions of people, and we'd probably make the same call in their position.

Engram works the other way around: nothing is saved unless you or your AI saves it, and once you've decided to save something, there's no second filter deciding whether you should have. The control lives at the other end — you choose what goes in and what comes back out, down to a single message. And separately, if what you need stored is an actual secret, Engram's zero-knowledge vault encrypts values on your machine before they ever reach us — that one is for the CLI and first-party API, not for AI connectors, by design.

Where Claude's memory is better

Four places, and none of them are small:

• It is already on. No account, no connector, no OAuth flow, no decision. Default-on beats better-but-optional more often than product people like to admit.
• It saves without being asked.Claude writes topics as you chat. Engram is a tool the model has to choose to call — and models don't always choose well. We ship a skill with the Claude Code plugin precisely to make that reliable, but it is a real difference and it's the main reason a new Engram user's first week can feel quieter than they expected.
• Short files are easier to correct.If Claude concluded something wrong about you, you edit one file and it's fixed everywhere. A verbatim archive has no equivalent gesture — you can delete, but you can't retroactively change what was said, which is the point and also occasionally the inconvenience.
• It costs nothing and always will.It's part of the product you're already paying for, or not paying for.

There's also a real availability caveat worth knowing if you're on a managed plan: Anthropic notes memory isn't available to Enterprise organizations under HIPAA, public-sector, or custom data-retention agreements, and on Team and Enterprise it's off until an owner turns it on.

But Claude has memory import and export now

It does, and it's a genuinely good thing that a major vendor shipped it — Anthropic documents it as a way to move memory between AI products. Read the fine print before you count on it, though:

• What exports is the memory entries — a plain-text list of things Claude concluded. Not your conversations. The transcripts stay behind.
• Anthropic calls import experimental and in active development, and says Claude may not always successfully incorporate what you paste in.
• It's a copy, not a sync. Paste your memory into another product today and the two diverge from tomorrow.
• Because memory is scoped to work topics, Anthropic warns imported personal details may simply not be retained.

Engram is the other shape: one live store that every tool reads and writes as you go, so there's nothing to migrate because nothing is ever in only one place. And the door out is the same door everyone else uses — the public API, which is how a customer built an Obsidian knowledge graph on top of their own memories without asking us first.

They are not really competing

The useful mental model: Claude's memory is a profile. Engram is an archive.

A profile should be short, opinionated, and loaded into every conversation — that's what makes Claude stop asking what stack you use. An archive should be complete, unedited, and searched on demand — that's what lets you ask “what exactly did we decide about the auth flow in June, and why?” and get the sentence you actually wrote.

Running both is the normal configuration, not a compromise. Let Claude keep its topic files; connect Engram so the conversations underneath them survive, and so ChatGPT and Cursor and your terminal can see them too. If you already have history worth keeping, both ChatGPT and Claude will hand you an export today, and Engram will ingest it — twice, if you like, without duplicating anything.

Try it in two minutes

The claim is testable, which is the best thing about it. Connect Engram free — 10,000 messages, no credit card — and save one real conversation. Then open a different AI entirely and ask for a specific detail from it: not the gist, the detail. A number, a name, the exact phrasing.

Ask Claude's memory for the same detail. Whatever comes back is the honest answer to this comparison, and it takes about as long as reading this page.