Rewind Shut Down. The Interesting Question Is Why Total Recall Was the Wrong Shape

The short answer

Rewind AI's Mac app shut down on 19 December 2025, after Meta acquired its successor company Limitless. The product worked: it recorded your screen and audio continuously, stored it locally, and let you search it. People who relied on it genuinely lost something.

But the shutdown is a good moment to be precise about what it was actually good at, because "AI that remembers" is now a crowded claim and Rewind's architecture answered a narrower question than its marketing implied. Recording everything gives you excellent retrieval. It does not give you understanding. Those are different problems, they need different data structures, and conflating them is why a lot of memory features feel impressive in a demo and useless on a Tuesday.

What Rewind was actually solving

The core insight was sound and slightly uncomfortable: you cannot know in advance what will matter later. Any system that asks you to tag, file or save is a system you will not use consistently, because the moment of capture is exactly when the information seems unimportant.

Rewind's answer was to remove the decision entirely. Capture the screen, capture the audio, OCR it, transcribe it, index it, store it on the device. Then search after the fact, when you do know what matters.

For a specific class of problem this is close to perfect:

• "What was that library someone mentioned on a call three weeks ago?"
• "I saw a chart with the right number in it. Which document?"
• "What did we actually agree in that meeting?"

These are lookup problems. You know a thing existed, you experienced it, and you need to find it again. A time-indexed recording with full-text search over OCR and transcripts is a very good answer, and the local-first storage made it defensible in a way a cloud version would not have been.

Where the shape runs out

The trouble starts when you want the assistant to use what it knows rather than hand it back to you.

Consider a question like "who should I loop in on the payments migration?" There is no single frame of video that answers it. The answer lives in a pattern spread across dozens of sessions: who has commented on payments code, who was in the architecture review, who raised the compliance question, who has gone quiet since. A recording contains all of that evidence and none of that structure. You can search it; you cannot reason over it.

This is the gap between a log and a model.

A log is append-only, high-fidelity, and undifferentiated. Every second has the same status as every other second. It is excellent for reconstruction and terrible for inference, because nothing in it has been resolved into an entity. "Priya" in a Slack screenshot, "P." in your notes, and a voice on a call are three unrelated blobs of pixels and audio to a recording. They are one person to you.

A model is the opposite: lossy, structured, opinionated. It says there is a person called Priya, she has a role, she is attached to these projects, she raised these concerns, and the last meaningful interaction was on a date. It throws away almost everything a recording keeps. What it keeps, it can reason over.

The storage tell

There is a practical signal that these are different problems: the storage profiles are not remotely comparable.

Continuous screen recording is video-scale. Even heavily compressed, it is tens of gigabytes a month, and that cost buys you fidelity you will use perhaps a dozen times a year. A structured world model of the people, projects and decisions in your working life is text-scale — kilobytes, not gigabytes — and it gets touched on every single interaction.

When one part of a system costs four orders of magnitude more than another and is used four orders of magnitude less often, that is usually a sign the two parts are not the same feature. Rewind bundled them because the recording was the only substrate it had. The structure had to be derived on demand, from pixels, at query time, which is both expensive and lossy in the wrong direction.

What a memory layer looks like if you start from the model

If you accept that understanding is the harder problem, the architecture inverts. Instead of capturing everything and deriving structure at query time, you capture the interaction and resolve structure as it arrives.

That is the approach behind Vyra's memory, and it has two layers rather than one:

Episodic memory is the log half, but scoped. Every interaction with the assistant is indexed for both keyword search (FTS5) and semantic search (sentence-transformer vectors), so "that thing about rate limiting" finds the right conversation even if you used none of the original words. It is a log, but a log of a conversation rather than a log of a screen — orders of magnitude smaller, and already textual.

The world model is the structured half. People, projects and knowledge exist as explicit objects with relationships, roles, milestones and blockers. When you mention that Priya is now leading the migration, that does not become a sentence in a transcript; it updates a field on an entity. The next time the goal engine reasons about the migration, it reasons with that fact rather than searching for it.

The two layers need each other. The world model without episodic memory is a database with no provenance — it asserts things and cannot show you where they came from. Episodic memory without a world model is Rewind's problem in miniature: searchable, not usable.

Consolidation is the part nobody demos

The piece that makes this work over months rather than weeks is unglamorous: something has to periodically go back over the episodic layer, decide what has proven durable, and fold it into the model.

Vyra runs this nightly, which is a deliberate borrowing from how memory consolidation is described in sleep research — not because the analogy is rigorous, but because the engineering problem has the same shape. Recent detail is cheap to keep and expensive to keep forever. Importance decays. Some of it turns out to matter and should be promoted into structure; most of it should be compressed and left where it is.

Without that step, either the model goes stale or the log grows until retrieval quality collapses. Both failure modes are slow, which is why they rarely show up in a product review and always show up in month four. There is more detail on how this differs from retrieval-augmented generation in agent memory vs RAG.

Being honest about what was lost

None of this replaces what Rewind did well, and it would be dishonest to pretend otherwise.

If your actual need is "find the window I had open in April", a structured memory layer cannot help you. It never saw the window. That is a genuine capability gap, not a design improvement, and if visual recall is what you miss then a dedicated screen-capture tool — Screenpipe and several others are active in this space — is the honest recommendation over any assistant, including ours.

What is worth separating is the two things people conflated when they praised Rewind. Some loved the timeline. Many more, judging by how people wrote about it, loved not having to re-explain themselves. Only the second of those requires recording your screen, and it was always the cheaper half to build.

We wrote up how Vyra compares in detail, including where it falls short, on Vyra as a Rewind AI alternative.

Common questions about Rewind's shutdown

Why did Rewind AI shut down?

Meta acquired Limitless, the company Rewind had become, and the Rewind Mac app was sunset on 19 December 2025. Pendant sales ended around the same period. The shutdown was a consequence of the acquisition rather than a failure of the product itself — it had a committed user base at the point it closed.

Is there a direct replacement for Rewind AI?

For continuous screen recording with searchable playback, the closest active options are dedicated screen-capture tools such as Screenpipe rather than general AI assistants. Assistants that advertise "memory" almost always mean structured recall of conversations, which is a different capability. Check which one a product actually offers before switching.

Was Rewind's local-first storage meaningful?

Yes, and it deserves credit. Continuously recording a person's screen is about as sensitive as software gets, and keeping that data on-device rather than shipping it to a server was the right call. The lesson worth carrying forward is that local storage makes an enormous capture surface defensible; it does not make it small. Reducing what you capture in the first place is the stronger privacy position, which is the argument in local-first AI and the privacy question.

Does an assistant need to record my screen to remember my work?

No — and this is the central claim of this piece. Recording is one way to acquire memory, and it optimises for reconstructing moments. If what you want is an assistant that knows who your collaborators are and what your projects need, the structured approach gets there with a fraction of the captured data, because it stores conclusions rather than raw evidence.

---

Vyra's memory is built model-first: episodic search over interactions, plus a structured world model of the people, projects and knowledge in your working life, consolidated nightly. It does not record your screen. If that trade sounds right, join the waitlist — closed alpha is running now, with a Founders Beta ahead of public launch.