Pieces for Developers

The trait that decides it is recall of your own week. Ask what you were working on Tuesday afternoon, or which article you had open when you made a decision, and you get an answer, which is a genuinely different thing from a tool that knows your codebase. Anybody who writes weekly updates or reconstructs a bug from memory will feel the difference on day two.

The cost is a recorder running all day, and some people will never be comfortable with that. Pick it if the answer to that is fine. Pick Letta if you would rather the memory belonged to an agent than to a log of you.

The trust requirement here is unusually large. A continuously running background process observes your activity across applications and builds a durable record, and the source is closed, so what is captured, what is discarded and what is indexed are all matters of vendor assertion rather than inspection. Local storage narrows the exposure; it does not make the capture legible.

Ask for a written statement of what is excluded, particularly password managers and terminals. What it does right: the model is chosen per question across several vendors, so no single provider sees the whole picture of your work.

1. Capture is continuous and passive, so the corpus is a timeline rather than a curated set, which changes the retrieval problem from relevance to recency-weighted relevance. 2. The query surface is stated as four dimensions: time, topic, person and tool, implying entity extraction over captured content. 3. That extraction step is where accuracy is won or lost, and it is not documented.

No benchmark and no recall figure accompany the product. For a memory system the interesting number is what fraction of correct answers are retrievable, and it is not published. The exposure of that memory to outside clients as a served interface is the more reusable idea here.

The packaging is the thesis. Starting at $18.99 with a seven-day trial and no free tier means every user is a paying decision within a week, which is either confidence or desperation, and here it looks like confidence. The reason is the asset: months of accumulated personal history are not exportable into a competitor, so churn falls sharply after the third month.

Moat: genuine and per-user, which is unusual in this category. Likely acquirer: an editor vendor that wants memory it cannot build without a desktop presence. Position: small, durable business, more likely acquired than scaled.

Sixty people on the enterprise plan is $1,379 a month, which is not the obstacle. The obstacle is that a plan named for enterprises should come with the things enterprises ask for, and I cannot find published detail on directory integration, provisioning, or an administrative view of what has been captured on each machine.

Without that, I am authorising sixty independent recording devices with no central control and no way to answer a subject access request. It has no role in our pipeline either. Onboarding is a download. Approved with conditions: a written administrative and deletion story first, then a twelve-person pilot.

Closed source and no repository, so I cannot read it, patch it or survive its vendor, and that is the ceiling on every score I give it. There is no configuration file to version and no key of my own to supply, which means the model routing is theirs to change.

And yet. The data stays on the device, local models are among the options, and it publishes an MCP server, so my own clients can query that memory rather than being locked out of it. A closed tool exposing its state over an open protocol is doing the one thing that makes it survivable. Grudging, but noted.