Letta

The trait you feel daily is continuity. You come back on Thursday and it still knows what you decided on Monday, which removes the ritual of pasting the same background into a fresh chat forever. Because the same agent answers in a terminal, a desktop app, a browser tab or a chat channel, it also stops mattering where you happen to be sitting.

What you give up is editing horsepower on hard repository work. Pick it when the relationship with the agent is the product. Pick Claude Code when the diffs are the product.

Start with the thing nobody mentions. The well-known repository has become a project landing page; the code that runs moved to a separate repository. Anyone auditing this by star count or issue history is auditing a signpost. Contributions, security reports and blame history all fragment across that boundary, and newcomers file issues in the wrong place for months after such a move.

Check which repository a release actually came from before you pin anything. Done right: it ships a self-hosted application server, so the accumulated memory can stay on hardware you control instead of theirs.

1. The agent is stateful by construction: its memory is held by a server and outlives any single conversation, which is the inheritance from the MemGPT line of work. 2. Because state is central, the terminal interface, desktop applications, browser client and chat channels are all views of one agent, not separate instances. 3. Memory is described as improving over time, implying a revision step rather than pure accumulation.

No benchmark is published for retrieval quality or memory accuracy, so the improvement claim is asserted. Measuring memory is hard, which is precisely why it needs measuring.

The packaging is smart. Three agents free is enough to form a habit and not enough to run anything real, so the upgrade moment arrives after the switching cost exists rather than before. There is also a usage-based interface plan alongside the personal tier, which means the company is testing whether it sells to people or to applications.

Moat: accumulated memory is genuine lock-in, the best kind in this category, because exporting a relationship is harder than exporting a repository. Likely acquirer: a model lab that wants persistent state it does not have. Position: long, with execution risk on packaging.

Twenty dollars a seat means $1,200 a month for sixty engineers, which is inside the noise on my tooling line. The blocker is not price. This product accumulates a persistent record of what each engineer works on, and no published material tells me where that record lives, how long it is kept or who at the vendor can read it.

There is no unattended mode, so it never becomes part of our pipeline, and no directory integration is described. Onboarding is a day. Approved with conditions: the self-hosted server, a retention answer in writing, and no customer data in memory.

Apache-2.0 and a global npm install gets me the client without a signup wall, and there is a TypeScript SDK if I want to drive the thing from my own code rather than its interface. Twenty-four thousand stars means a fork would find contributors on day one. So far, good.

Then the doors: it is neither an MCP client nor an MCP server, so every tool is its own integration, and local models are not supported, so the memory may live on my box but the thinking never does. For a project whose whole pitch is owning state, outsourcing every token is an odd place to land.