Letta Agent launches persistent digital coworkers with Slack, Discord, and BYOK state

• Letta shipped a desktop agent product that it pitches as a persistent digital coworker: teachable, customizable, shared across a team, and able to run locally or on remote machines, according to Letta's launch post.
• The core mechanic is long-term state. As Letta's memory thread describes it, each agent rewrites its own skills, prompts, and reference files over time.
• Letta also shipped channel continuity. Letta's channels post says the same agent identity can show up in Slack, Discord, Telegram, and custom channels.
• Model choice is deliberately loose. In Letta's BYOK post, Letta says users can connect an existing ChatGPT subscription or API keys and port the same agent state across model upgrades.
• The most concrete early demo came from Sarah Wooders' Slack screenshot, where a fully local finance agent installed a web-search mod from Slack and immediately used Exa with an existing API key.

You can see the product pitch in Letta's launch post, the memory behavior in the learning thread, and the channel model in the Slack and Discord thread. The sharpest extra detail is Sarah Wooders' demo, where the agent installs a mod, registers a new tool, and tests Exa-backed search from inside Slack. Even the surrounding practitioner chatter is converging on the same problem shape: Matt Pocock's reply treats folder-scoped agent context as a live experiment, while another Pocock reply separates rarely read wiki state from constantly read skills files.

Letta's main bet is that the agent should accumulate its own operating context instead of starting clean every session.

The thread breaks that state into editable artifacts:

• skills
• prompts
• reference files such as a wiki

That is a more opinionated design than plain chat history. Letta's memory thread says the agent rewrites those files through use, which makes the unit of memory something closer to a maintainable workspace than a transcript.

Letta is framing the agent as a shared team actor, not a one-user sidebar.

According to the launch post and the channels thread, the same agent can:

• work with a whole team
• run on local or remote machines
• appear in Slack, Discord, Telegram, or custom channels

That combination pushes the product toward a durable identity. The same stateful agent is meant to survive both model changes and surface changes, instead of being recreated per app.

Letta's model layer is intentionally swappable.

Letta's BYOK post says users can attach an existing ChatGPT subscription or API keys, then port the agent with its configuration and state to newer model releases. The launch thread also says the app is available on macOS, Linux, and Windows, via Letta's launch post.

The launch thread talks about persistence. Sarah Wooders' demo shows what that looks like when an agent gains a new capability midstream.

The screenshot shows a few implementation details that the announcement thread does not spell out:

• the agent installed @letta-ai/web-search from Slack
• the mod registered a web_search tool
• the tool exposed multiple providers: Exa, Tavily, Parallel, and Perplexity
• the agent found EXA_API_KEY already in its environment
• it immediately tested a San Francisco weather query and returned results

That makes the product feel less like a fixed assistant and more like a persistent runtime that can extend itself while staying inside the same team thread. The adjacent practitioner discussion points in the same direction: Matt Pocock's reply describes folder-scoped agent context as something he plans to keep experimenting with, and his follow-up argues that a wiki is lower-stakes than a constantly read skills file.