Letta Agent is the easiest way to add persistent, digital coworkers to your team. Each agent:
- is teachable and customizable
- never forgets
- works with your whole team (not just you)
- runs locally & on remote machines
Available on MacOS/Linux/Windows
3/ Model Agnostic + BYOK
Connect your existing ChatGPT subscription or LLM API keys. Port your agent (including all its configuration and state) to the latest model release.
2/ Works where you work
Bring your agent to where your team works.
Chat with the same agent on Discord, Slack, Telegram, or through custom channels - just like you would a human coworker.
1/ Learning & memory
Each agent becomes unique through its own experience, and what you & your team teach it.
As you work with your Letta Agent, it will rewrite its skills, prompts, and other reference files (e.g. a wiki) over time.
Letta Agent is the easiest way to add persistent, digital coworkers to your team. Each agent:
- is teachable and customizable
- never forgets
- works with your whole team (not just you)
- runs locally & on remote machines
Available on MacOS/Linux/Windows
Mods in Letta are pretty cool - agents can self-expand their capabilities.
Today, I had our fully local finance agent (running on @Railway) install a mod to support web search with @ExaAILabs -- all via slack
Agents are now very good at writing code. Agents are also composed of code (the harness). This means that agents can self-adapt through rewriting their harness - now supported in Letta Code through *mods*.
Mods are very similar to extensions in @badlogicgames's Pi harness. You
Agents can now not only rewrite their memory, but also rewrite the execution code they run in (the harness) through Mods.
Mods (inspired by Pi's extension system) allow for harness-level changes like:
- modifying context
- injecting custom tools
- customizing the statusline
Harness is memory.
My current running example: "please use uv not pip".
This can get written into system-level memory (context-as-memory), or if the agent has access to mutate its own harness, it can be written into a pre-tool use hook (block commands w/ pip).
Agents can now not only rewrite their memory, but also rewrite the execution code they run in (the harness) through Mods.
Mods (inspired by Pi's extension system) allow for harness-level changes like:
- modifying context
- injecting custom tools
- customizing the statusline
Agents can now not only rewrite their memory, but also rewrite the execution code they run in (the harness) through Mods.
Mods (inspired by Pi's extension system) allow for harness-level changes like:
- modifying context
- injecting custom tools
- customizing the statusline
External mods can also be installed through npm or git with: `letta install <source>`
We're listing community and official mods at letta.com/agent/mods/
Very exciting times in research for token-space continual learning, with ideas like dreaming finally hitting the mainstream.
One promising research direction we're excited to share more on is memory models, models that are trained explicitly for sleep-time compute / dreaming.
