Letta’s recent blog post on “sleep-time compute” has sparked substantial discussion in the agentic AI community, and it’s directly shaping the trajectory of the Sanctum project. We’ve been evaluating Letta’s framework in parallel with our ongoing work, and it’s worth highlighting how their approach is informing—and diverging from—what we’re building here.

Letta’s Sleep-Time Compute: A Quick Overview

The “sleep-time compute” concept introduced by Letta proposes that agents benefit from scheduled downtime, during which they process chat histories, reorganize context, and essentially “refresh” their internal models. This is loosely inspired by biological sleep cycles, and it’s an exciting area for both research and production-grade deployments.

For those running Letta, the new sleep-time agent feature can be found in the creation modal. This makes it accessible for experimentation right now.

How Sanctum Is Extending the Concept

While Letta’s implementation is promising, our initial testing has surfaced several limitations for the kind of multi-user, persistent environments we’re building with Sanctum:

• Configurability: The summarization and reset routines in Letta’s sleep-time agents are not yet fully customizable or directly addressable.
• Multi-User Dynamics: Letta’s default configuration is better suited to single-user deployments. In scenarios where multiple users interact with an agent, there’s a tendency for the agent’s personality to “average out,” losing individual nuance over time.

In contrast, Sanctum’s ongoing development includes a more robust, configurable approach to background compute and context refresh. Our aim is to support both deep historical processing (digesting years of chat logs) and regular personality resets—crucial for agents serving communities or switching between multiple user profiles (see: @myGirlMonday).

We’re also advancing threading logic, enabling Sanctum-based agents to maintain distinct conversational threads and minimize unwanted drift toward generic responses. This is especially relevant for collaborative deployments, where agent identity and contextual memory must be preserved across varied interactions.

Community Updates & What’s Next

We’re seeing a lot of innovation in the self-hosted agent space this week, and it’s encouraging to witness these ideas being tested and shared so rapidly. As always, we’re iterating in public and welcome contributions, testing, and feedback from the broader community.

If you’re experimenting with Letta, we encourage you to try out the new sleep-time agent, share your observations, and let us know where you see room for improvement. Meanwhile, Sanctum’s team is pushing forward with deeper configurability, improved context management, and support for persistent, multi-user agents.

Stay tuned for more technical breakdowns, release notes, and collaboration opportunities in the coming weeks.