News Hub

Letta released an SDK for developing stateful AI agents.

Evaluating Memory in Production Agents.

Letta proposes Trajectory as a standard format for storing and analyzing agent experience data.

Letta develops memory models enabling agents to learn over time.

Letta introduces Mods, allowing agents to self-improve through harness-level adaptation.

Red-teaming the Context Constitution audits models as experiential AI agents.

Letta released a code application for managing AI agents.

Letta provides orchestration for Claude Code and Codex agents.

Letta introduced context repositories using Git-based memory for coding agents.

Letta introduced shared agent memory across concurrent experiences called Conversations.

Letta Code is a coding agent that uses memory-first architecture for code generation and execution.

Continual learning in token space enables agents to learn and update knowledge within context windows.

Skill learning extends continual learning capabilities to command-line interface agents.

Tool calling can be implemented programmatically with any large language model.

Context-Bench is a benchmark for evaluating how LLMs handle agentic context engineering.

Letta Evals provides evaluation methods for agents that learn and adapt over time.

Letta redesigned its agent loop incorporating lessons from ReAct, MemGPT, and Claude Code.

Letta introduces Claude Sonnet 4.5 and Memory Omni-Tool integration.

Letta introduced Recovery-Bench, a benchmark for evaluating large language models.

Study examines whether a filesystem is sufficient for AI agent memory systems.