DoneThat vs No AI memory

The short answer

Keeping no persistent memory of your work is the safest option available. Nothing recorded is nothing to leak, nothing to maintain, and nothing to regret. Forgetting is genuinely a feature.

The cost is paid in a currency that is easy to miss: you re-supply context by hand, every time, to every tool and every person who needs it — including yourself in six months.

The case for keeping nothing

Nothing recorded is nothing at risk. No archive to breach, no dataset to be subpoenaed, no vendor holding a searchable record of years of your working life. For anyone handling sensitive material, this is not paranoia, it is the correct default.

Forgetting is useful. Old context is often wrong context. A system that remembers everything can keep surfacing a project you abandoned or an approach you deliberately moved on from.

No maintenance. No archive to curate, migrate between tools, or keep paying for.

Plenty of work does not refer back. Short-cycle work, support queues, anything where last month has no bearing on today — for that, a memory layer is a cost with no return.

What it costs

The cost is re-supplying context, and it is mostly invisible because it is spread thinly across every day.

The clearest symptom is with AI assistants. Every conversation starts from zero. You explain the project, the stack, the constraints, the thing you already decided against — and then you do it again next week, because the assistant that helped you has no idea it did. The work of getting an assistant up to speed is not small, and paying it repeatedly is the tax of no memory.

The second cost is re-derivation. Without a record, recurring work gets solved from scratch rather than recalled. The decision you made last spring, and the reason behind it, get reconstructed at full price — or, more often, differently, because you no longer remember the reason and quietly optimise for something else.

The third is that handover and review run on recollection. Retrospectives, performance conversations, project post-mortems: all of them ask what happened, and memory answers with a flattering summary.

Chat history is not the same thing

Worth separating, because people assume it covers this. An assistant's chat history remembers your conversations. It has no idea what you did between them — which is nearly everything, and usually the part that matters. Memory of what you said about your work is not memory of your work.

Why the timing matters

Memory has a peculiar property: the value compounds and the start date is unrecoverable.

A memory layer with two weeks in it is a novelty. One with two years answers what you decided last spring, how long that kind of project actually takes, and what you were doing the month everything went wrong. Nothing recovers the period before you turned it on, so deciding later is strictly worse than deciding now — which is an argument for starting even if you are unsure, and the closest thing to a real reason to act on this page rather than bookmark it.

The honest risk

An archive exists, which is both the point and the exposure. DoneThat's answer is to control what enters it rather than to keep it thin:

• Screenshots are never stored. Processed in real time, immediately deleted; only the derived structured activity persists.
• Excluded apps are redacted on your machine, before anything is sent for analysis.
• Bring your own LLM — your Gemini or OpenAI-compatible model, so raw activity never reaches our servers. Keys are encrypted locally.
• Open source desktop app, so the capture path can be inspected rather than trusted.
• Visibility is set per summary by the person it belongs to, and private is fully supported.

If those controls do not satisfy you, keeping nothing remains the safer choice, and that is a coherent position rather than a failure to understand the product.

Where no AI memory is the better choice

• You handle material that should not persist anywhere.
• Your work is short-cycle and rarely refers back.
• You do not use AI assistants for work, so the re-explaining cost never arises.
• You already have a system that works — notes, a wiki, a good memory — and it is enough.
• You would not use it. An archive nobody queries is pure cost.

Where DoneThat is the better choice

• You keep re-explaining your context to an assistant that helped you last week.
• You want to ask questions about your own past work and get an answer better than a guess.
• You want AI tools to reason from a real record — MCP and an API expose it as long-term memory.
• Handover, reviews and retrospectives matter and you would rather not run them on recollection.
• You want coaching that knows your history. Don nudges you onto goals, blocks calendar time, flags low-value work while you can still fix it, checks your mood, reminds you of meetings, suggests breaks.

How we compared these

There is nothing to cite: no AI memory is the absence of a system. Claims about DoneThat are first-hand — we build one of the two options here, which is the bias to read this with.

The strongest argument against us is the simplest one, and we have not tried to talk around it: an archive that does not exist cannot be misused. Everything above is about whether the controls make the trade worth it for you, not about whether the trade exists.

Questions people ask about No AI memory

Do I need AI long-term memory?

Only if you repeat context. The clearest signal is how often you find yourself re-explaining your project, your stack or your constraints to an assistant that helped you with the same thing last week. If that never happens — short-cycle work, little reference back — no memory layer is the cheaper and safer choice.

What is the risk of keeping a work archive?

It exists, which is the whole point and the whole risk. DoneThat's answer is to control what enters it rather than to keep it thin: screenshots are never stored, excluded apps are redacted on your machine before anything is sent, you can use your own LLM so raw activity never reaches our servers, and the desktop app is open source.

Isn't a chat history enough memory?

It is memory of your conversations, not of your work. It knows what you told an assistant and nothing about what you actually did between those conversations — which is most of it, and usually the part that matters.

Why start now rather than later?

Because the value is in the depth. A memory layer with two weeks in it is a novelty; one with two years is the thing that answers what you decided last spring and why. Nothing recovers the period before you started, so a later start is permanently worse than an earlier one.