Letta

Letta Agent is a platform for building AI agents that don't reset between conversations. Each agent maintains its own memory, identity, and skills across every interaction, improving continuously without requiring model retraining. It grew out of MemGPT research at UC Berkeley's Sky Computing Lab, which makes its approach to AI memory and context unusually rigorous compared to most agent tools.

The core idea is persistence. Most AI assistants start fresh every session. Letta agents accumulate knowledge over time, storing context, skills, and prompts in a git-tracked history so you can audit exactly how an agent has evolved.

Key capabilities:

• Persistent memory that survives across sessions, model switches, and device changes, with no memory loss when transferring between models
• Scheduled tasks that agents can set themselves or that you configure, with full run history available for review
• Multi-channel communication through Slack, Discord, Telegram, or custom integrations, so your team can teach and direct agents without leaving existing workflows
• Remote device support lets agents operate across multiple machines, sandboxes, or VMs simultaneously
• Bring your own model compatibility with OpenAI Codex, Z.AI, and other providers using your own API keys or existing subscription plans
• Custom skills and tools with harness modifications, giving each agent a distinct identity and area of expertise

Letta Agent is built for teams that want agents to get genuinely better with use, not just respond to prompts. If you're comparing it to tools like Mem0 for memory or LobeChat for multi-model chat, Letta's differentiator is the combination of long-term learning, proactive scheduling, and full auditability of how agents change over time. It runs locally or in the cloud, and connecting your own API keys is free.