ChatGPT Computer History vs. Kepler: recorded memory vs. the memory you choose
Is Kepler an alternative to ChatGPT's Computer History? Both give AI a memory, but one records everything you do and the other keeps only what you choose. Here's the difference, the research on why undifferentiated context makes answers worse, and when each one fits.
OpenAI shipped Computer History, and it’s the most interesting bet on AI memory anyone has made this year. Turn it on, and ChatGPT starts building a picture of what you’ve been doing on your Mac, which apps, which sites, what you were working on, so you can ask it “what was I in the middle of yesterday?” and get a real answer.
We build Kepler, which is also a memory layer, so the question we keep getting is the obvious one: is Kepler an alternative to ChatGPT’s Computer History?
Sort of. They both give AI a memory, so they’ll show up in the same conversation. But they answer different questions. Computer History answers what did I do? Kepler answers what did I decide was worth keeping? Those are two different questions, and they need two different kinds of memory.
There’s a familiar version of this split: it’s the same one your own mind makes. Computer History is like implicit memory, the passive background record of everything you did. Kepler is explicit memory, the things you deliberately chose to keep.
What Computer History actually does
Computer History is a well-built feature, and it’s worth understanding exactly what it does. It turns your recent activity across apps and websites into summaries and timelines that ChatGPT and Codex can read. It works through macOS accessibility APIs, capturing an interaction-event stream: clicks, typing, keyboard shortcuts, app switches. That stream gets periodically summarised into text-based memories that live locally on your Mac as plain Markdown files. Crucially, it records that activity indiscriminately: there’s no discretion about what’s worth keeping and what isn’t, because a recorder can’t make that call.
OpenAI has been careful with the edges. It’s off by default and every person has to opt in individually. You can exclude specific apps and sites, or flip it around and allow only the ones you whitelist. You can pause collection from the menu bar and clear history item by item or by timeframe. It explicitly excludes screenshots, microphone and audio, and private-mode browsing. The raw event files are processed on OpenAI’s servers to generate the memories, then retained no longer than they need to be: 48 hours for temporary event files, and OpenAI says it doesn’t keep them after processing unless legally required.
The practical limits are worth knowing. It needs the macOS desktop app and Memories turned on, and it isn’t available in the EEA, Switzerland or the UK. There’s also a subtler limit worth naming: capturing this much context is genuinely powerful, but it cuts both ways. The same firehose that can reconstruct your week also floods your memory with context that was never worth remembering, and that noise oversaturates the very memory it’s meant to sharpen. More captured is not the same as more useful.
What Kepler does instead
Kepler is a shared memory layer for the web. You save links into a space, a project folder for context, using the web extension or the mobile share sheet. Kepler pulls the usable context out of each save, and your highlights on the parts that actually matter become relevance signal. Then any AI tool you use can read that space over MCP: ChatGPT, Claude, and others.
Kepler isn’t an AI model and it doesn’t replace ChatGPT. It’s the layer underneath, the thing that gives the model you already pay for something specific and yours to reason over.
The part that matters for this comparison: nothing enters Kepler unless you put it there. Connecting a source grants permission; it doesn’t start a crawl. There’s no background indexing of your accounts, no observation of what you’re doing, no sweep of your history. One save at a time, on purpose.
Recording is cheap. Choosing is the signal.
Here’s the thing an event stream can’t give you: your judgement.
If a system watches everything, every item arrives with the same weight. The article you read closely and the tab you opened by accident and closed in four seconds look identical to a summariser. It can guess at importance from dwell time and repetition, but it’s guessing, because the one signal that would settle it, you thought this was worth keeping, was never captured.
When you save a link to a Kepler space and highlight three lines in it, you’ve done something a recorder structurally cannot do. You’ve said: this one, and specifically this part of it. Weeks later, that’s the difference between “here’s a timeline of your browsing” and “here’s what you concluded, and here’s what you flagged as still unresolved.”
That curation compounds. A year of saves and annotations is a retrieval graph that is genuinely yours, and it’s portable, which a device-local activity log isn’t.
Why capturing everything is usually more than you need
There’s a reasonable instinct that says: capture it all, sort it out later, storage is cheap. We think it’s the wrong default for memory, for three reasons.
Most of what you do isn’t worth remembering. A week of your clicks is overwhelmingly noise: Slack, email, the same six tabs, a rabbit hole you abandoned. Feeding all of it into memory doesn’t produce a better answer; it produces a more confident answer built on a worse ratio.
And this isn’t only intuition, it’s measurable. Stanford’s Lost in the Middle study found that models use information best at the start and end of a long context and routinely miss it in the middle, so simply having the right fact in memory doesn’t mean the model will actually use it. A Google study, Large Language Models Can Be Easily Distracted by Irrelevant Context, went further: adding a single irrelevant sentence to a problem the model would otherwise solve correctly measurably dropped its accuracy. And Chroma’s 2025 Context Rot report tested 18 frontier models and found every one degraded as input grew, sometimes losing 30 to 50 percent of its accuracy well before the context window was anywhere near full. Undifferentiated context doesn’t just cost more; it makes the answer worse.
The privacy surface scales with the capture surface. OpenAI has fenced Computer History carefully, and the controls are real. But the honest framing is that opt-in-then-exclude means you’re maintaining a blocklist against a system that is otherwise watching. Kepler’s default is the inverse: the system sees nothing until you hand it something. There’s no blocklist to maintain because there’s no sweep to fence off.
A record of activity isn’t a record of thinking. Computer History is very good at reconstructing the shape of your week. It’s not built to be the place your research on a topic lives, organised, annotated, and reusable six months from now. Different job.
Side by side
| ChatGPT Computer History | Kepler | |
|---|---|---|
| Capture model | Passive: observes app and website activity once enabled | Deliberate: you save each item on purpose |
| What it captures | Interaction events: clicks, typing, shortcuts, app switches | Links you choose, plus your highlights and notes on them |
| Relevance signal | Inferred from activity patterns | Supplied by you: annotation the AI can’t derive from content alone |
| Which AI can read it | ChatGPT and Codex | ChatGPT, Claude and other AI tools via MCP |
| Where it works | macOS desktop app; not available in the EEA, Switzerland or UK | Web extension and mobile share sheet |
| Default state | Off; opt in, then exclude apps and sites you don’t want seen | Empty; nothing is captured until you save it |
| Sharing | Personal to your machine | Shared spaces for teams, partners and collaborators |
| Best at | ”What was I working on?” for resuming and reconstructing your week | ”What did I conclude?” for research, synthesis and reusable project context |
These aren’t mutually exclusive
We’re not going to pretend Computer History is bad at its job. If you’re on a Mac, you live in ChatGPT, and you want to pick up an interrupted task, it will do something Kepler doesn’t try to do. Use it.
But the memory that survives is the memory you chose. It’s the same reason your explicit memories outlast the passive blur of an ordinary day: you decided they mattered. The moment your work spans more than one AI tool, more than one person, or more than a few weeks, an activity log on one machine stops being the right container. That’s the gap Kepler fills, and it fills it without watching you to do it.
Get started
Start a space, save the next ten things you actually want to keep, and connect it to whichever AI you already use. It’s free, and connecting over MCP is free and unlimited on every tier. It never costs AI credits.
Kepler helps you create context for AI from the links you save. It improves the AI tools you already use. It isn’t one itself.
Sources
- OpenAI, Computer History documentation: learn.chatgpt.com/docs/customization/computer-history
- Liu et al., Lost in the Middle: How Language Models Use Long Contexts (2023): arxiv.org/abs/2307.03172
- Shi et al., Large Language Models Can Be Easily Distracted by Irrelevant Context (ICML 2023): arxiv.org/abs/2302.00093
- Chroma, Context Rot: How Increasing Input Tokens Impacts LLM Performance (2025): trychroma.com/research/context-rot
