Khwan vs Letta
Letta runs the agent. Khwan runs nothing — keep your harness, keep your model, add memory over MCP.
Where Letta is the better choice
Letta is fully open source and self-hostable, and Letta Code is the #1 model-agnostic open-source coding harness on TerminalBench — ahead of the previous leader and comparable to harnesses built by the model providers themselves. It has skill learning, sleep-time compute, and git-style versioning of context. If you want a coding agent, that is a strong product and Khwan is not one.
The question is how much you have to adopt.
Letta runs the agent, which means moving onto their runtime. Khwan runs nothing: your existing harness stays, your model stays, and memory arrives over MCP in one line. If you already like Claude Code and want it to stop forgetting your corrections across projects, that is a smaller change than switching harness.
Inference on the read and write paths.
Letta's sleep-time compute has the agent reason about its context during idle periods — that is inference, and it is central to how the product learns. Khwan's /prepare and /record make no model calls; distillation is a single scheduled pass rather than continuous background reasoning — 8 brains and 245 stored turns for $0.0094 in a real production run.
One brain per project, deliberately isolated.
Khwan's unit is an isolated core — a separate brain per repository, client, or tenant, four of them on the free plan, with no blending between them. It is the shape consulting and agency work already has.
| Khwan | Letta | |
|---|---|---|
| Model calls on read/write | None | Sleep-time compute |
| Runs your agent | No — yours does | Yes |
| Adoption cost | One MCP line | New harness |
| Corrections as a typed event | Yes | Skill learning |
| Isolated brains, free plan | Four | Limited agents |
| Open source | No | Yes — fully |
| Coding-harness benchmark | n/a | #1 OSS on TerminalBench |
| Self-hosting | Not proven | Yes |
Letta details verified 2026-08-27 from letta.com, Letta Code, Letta pricing. Products change — if something here is out of date, tell us and we will fix it.
Take Letta if you want an open-source agent runtime, and Letta Code if you want a coding agent that learns. Take Khwan if you already have an agent you like and only want to change what it remembers.
Other comparisons
See all →Both give an agent memory that survives the session. Mem0 runs a model to maintain it; Khwan's read and write paths make no model calls at all.
vs ZepZep keeps facts accurate over time, at enterprise scale. Khwan captures the moment a person says the agent was wrong. Those are different problems.