Devin Gets a Memory That “Dreams”, and Hark Pro Uses Your Computer for You

Two agent launches this week tackle two different gaps. Cognition has given its Devin coding agent a way to remember what it learns about you between sessions, and to tidy those memories each night [1]. Hark, a startup less than a year old, has released Hark Pro, a free assistant that operates your computer on your behalf [2].

Both launches are confirmed. How well they work is, so far, the companies’ own account.

What has Cognition added to Devin?

Devin is an AI agent that writes and fixes code. On 5 October Cognition introduced Memory and Dreaming [1].

Memory lets Devin keep lessons across sessions: your preferences, corrections you made, and things it learned about your projects. Cognition stresses that these are “not summaries of sessions” but short notes, each linked back to the session where the lesson came from [1].

The notes live in what Cognition calls a Memory Drive: a Git repository (the version-control system programmers use to track changes) full of plain text files. A short index file, MEMORY.md, holds general preferences and points to the other notes. Devin reads that index at the start of each session and looks up other notes only when they seem relevant, rather than loading everything into its prompt [1]. When two sessions update memory at once, Git’s revision checks stop one silently overwriting the other [1].

Users can browse what Devin has stored under Customize → Memory. Memory is personal to each user within an organisation [1]. Cognition has also open-sourced the format, inviting other agent builders to adopt it [1].

What is “dreaming”?

Dreaming is a daily background session. Devin reviews past conversations alongside its stored notes and, according to Cognition [1]:

• merges notes that overlap;
• removes passing details (Cognition’s example is a note about which port a test server used one afternoon);
• looks for lessons it missed at the time;
• removes stale records that no session has used.

The name is a metaphor. As Cognition describes it, what changes is the set of notes, not the model itself.

That last step is worth a closer look. A note can go unused and still be true, or keep being used after it has become wrong. Cognition’s description defines “stale” by use, not by checking whether a fact still holds [1]. TSN’s earlier piece on AI memory argued that telling current facts from outdated ones is the hard part; Devin’s notes make that problem visible and editable, which is progress, but they do not make it go away.

What is Hark Pro?

Hark was founded less than a year ago by serial founder Brett Adcock. On 6 October it widely released Hark Pro, which is free, with a paid tier for heavy users, TechCrunch reports [2].

Hark Pro runs on a model trained specifically for computer use: rather than only answering questions, it operates websites and apps for you. Users connect it to their email, calendar, files and payment cards, and it suggests and carries out tasks [2].

One design choice stands out. Hark Pro shows the agent browsing in a small window as it works, which most assistants do not. Hark’s design lead, Abidur Chowdhury, says this is meant to build trust that the agent is “doing the right thing” [2].

Hark also pitches itself on privacy. “We’re not here to like sell you ads and steal your data,” Chowdhury told TechCrunch [2]. That is a company statement, not an audited policy. On rivals, he says: “It might not be faster than Muse, but it will give you a better result, and it will work end-to-end” [2]. That too is Hark’s own claim, with no published test behind it.

TechCrunch raises one open question: a model trained for computer use may give up some of the broader knowledge of a general frontier model [2]. Hark also says it will release its own AI device in 2027 [2].

What links the two?

Both launches tackle trust. Devin’s memories sit in plain files you can read and correct. Hark’s agent works where you can see it. Both are ways of letting people check what an agent is doing instead of taking it on faith.

What this does not prove

• That Devin’s memory makes it more accurate. Cognition has not published before-and-after results [1].
• That dreaming removes wrong notes. It removes unused ones; that is not the same thing [1].
• That Hark Pro beats rivals. The comparison with Muse is Hark’s claim [2].
• That Hark handles data more safely than others. The privacy pitch is Hark’s own description; no independent review is cited [2].

The Bottom Line

Agents are getting notebooks and hands. Cognition’s Devin now keeps short, inspectable notes about how you work and tidies them nightly, in a format others can reuse [1]. Hark Pro will click through websites for you, in a window you can watch [2]. Both are useful steps towards agents people can check. Neither yet shows that the agent remembers the right things, or does the job better than its rivals.

Related on TSN: It Sounds Like It Remembers: The Memory AI Models Still Don’t Have

Sources

1. Cognition, “Memory and dreaming: how Devin learns from working with you,” Devin blog, 5 October 2026. https://devin.ai/blog/memory-and-dreaming
2. Tim Fernholz, “Hark releases an AI personal assistant with a focus on privacy,” TechCrunch, 6 October 2026, 19:22 BST (11:22 PDT). https://techcrunch.com/2026/10/06/hark-releases-an-ai-personal-assistant-with-a-focus-on-privacy/

Source note: Cognition’s open-source memory standard is linked from its post (cognition.ai/agent-memory-repo); TSN did not review the repository itself. Hark details come from TechCrunch’s report and interview; Hark’s own site was not used.