OpenHuman

The trait that decides it is where it lives. This is not another tab you have to remember to open; it reaches you in the messaging apps already on your phone, which is the difference between a personal agent you use and one you install twice and forget. Combined with a memory that builds up locally, it starts to feel like a thing that knows you.

That is also the reason to be careful, because you are handing it a lot. Pick it if a personal daemon appeals and you accept the exposure. Pick Letta for memory with a narrower blast radius.

The exposure compounds in three directions at once. It executes commands, it drives a browser with your sessions, and it accumulates a persistent local record of what you do. No container or virtual machine isolation is listed. The project labels itself preview, which is an honest disclosure and also an admission that none of those three surfaces has been hardened yet.

Do not connect it to accounts you cannot afford to lose while it carries that label. What it does right: the memory mirrors into a standard notes vault, so you can read and delete what it knows using a tool you already trust.

1. Context is retained by compressing observations into Markdown trees stored in SQLite, which means the summarisation step is a design decision with a lossy stage a reader can inspect. 2. Sub-agents run over checkpointed graphs, so a long task resumes from its last checkpoint rather than restarting. 3. Toolsets are partitioned by domain into browsing, coding and voice.

Nothing describes what the compression discards, which is the question that determines whether the memory degrades gracefully or confidently. Documentation is a hosted book with no benchmark, and the listing carries no verification date.

The revenue model is resale. The paid tier bundles managed inference, a commercial search interface and media generation, all three purchased upstream and marked up. That is a real revenue line with structurally thin margins, and it competes against the free path of pointing the same software at your own accounts, which the project itself documents.

Moat: the accumulated personal memory, which is genuine switching cost, and thirty-nine thousand stars of attention as the funnel. Likely pivot: away from resale toward a hosted service where the state lives with them. Position: interesting asset, unproven economics, too early for the cheque.

I do not get to the cost line on this one. A tool that ingests an employee's messages across personal and work channels and stores the result creates processing obligations we would have to document, disclose and defend, and it does so on hardware we do not administer. Legal would stop this before finance saw it.

There is no directory integration, no central audit trail, no retention control and no support agreement to attach obligations to. It also has no unattended pipeline role, so there is no upside to weigh against any of it. Not yet, and not on company machines.

GPL-3.0 is the licence I want on something this personal, because any fork has to stay readable and nobody can quietly close it. A local Ollama endpoint is supported, so the model can run on my own hardware and the whole loop stays behind the firewall, which for a system holding this much about me is not a preference but a requirement.

It speaks MCP as a client, so my servers extend it without waiting on anyone. There is no published install command at all, which means cloning and building, which is how I would have installed it regardless.