AI Potluck
Back to Gap Map Product / UX / Orchestration & agents

Letta

Letta
open source / Overall score: 3.7

Agent harness for long-lived agents that keep persistent memory and rewrite their own memory blocks, skills and prompts as they work. Letta Code runs as a terminal interface, a desktop app and an App Server that applications reach through the Letta Agent SDK, and its agents can run on schedules, call subagents and answer in Slack, Telegram or Discord. Letta, the team behind MemGPT, builds it.

Read against letta-ai/letta-code, where Letta's current source lives. The original letta-ai/letta repository now holds only a pointer to it, and the retired Python server sits on an archive branch.

Openness

5 high confidence
5.0
license
Apache-2.0(Letta Code harness, terminal interface and App Server)+MIT(bundled self-configuration skill)
source
public
core features withheld
no — the 2,450-path tree at 94145d7 carries the root Apache-2.0 LICENSE and one MIT skill LICENSE, with no ee/, enterprise/, commercial/ or proprietary/ path
commercial
Letta-Cloud

Letta Code is Apache-2.0 licensed, and its public source holds the harness, terminal interface and App Server, which run entirely on the user's own machine without a Letta account. Letta's paid plans sell hosted services beside it, such as cloud-stored agent state, remote sandboxes and model usage, rather than features cut from the code.

Adoption

3 medium confidence
3.0

Adoption is read from npm downloads of @letta-ai/letta-code over the trailing month, the package Letta's own instructions install. The letta package on PyPI is left out because its download history mixes the retired Python server with the current wrapper, and the desktop app is not counted.

Capability

4 medium confidence
4.0

Letta Code agents keep persistent memory, rewrite their own memory, skills and prompts, call subagents and run on schedules across several messaging channels. That is the same self-improving, long-running agent design as Hermes Agent, and like Hermes Agent it has no published benchmark result.

  • https://docs.letta.com/self-hosting/app-server recorded 2026-10-07

    The App Server page says it "lets you deploy the Letta agent harness as a service" and "can manage running multiple agents in parallel in a single process", reached through the Letta Agent SDK or a WebSocket protocol.

  • https://raw.githubusercontent.com/letta-ai/letta-code/main/README.md recorded 2026-10-07

    The README's feature table lists self-improvement and learning through memory blocks and skill learning, message search, MemFS, skills, subagents and multi-agent calls, messaging integrations, hooks, permissions, and crons and schedules; no benchmark figure.

Verified 2026-10-07, except some scores, still awaiting re-confirmation