Letta overview: persistent memory for local AI agents

Letta is presented as a framework for keeping an agent's knowledge across sessions. The speaker focuses on recurring development work: remembering test commands, generated files, project preferences and failed fixes so a coding assistant needs less repeated explanation.

The overview describes MemFS as a Git-backed memory file system that exposes knowledge as Markdown files. Some memory enters the system prompt; other material stays in files until needed. Version history lets users inspect edits and correct stale information. The speaker says agents can switch models during a session while retaining separate state, with local providers including Ollama, LM Studio and llama.cpp.

Hardware guidance covers 8GB and 16GB machines, then 24GB GPUs such as the RTX 3090 and 4090. These are broad examples rather than guaranteed requirements: model size, quantization and context length affect what fits. The reported memory benchmarks also concern specific tests, rather than proving general coding ability.

The speaker describes a terminal tool, desktop apps for macOS, Windows and Linux, and an app server for local or self-hosted agents. The speaker also describes optional cloud agents for state shared across machines, and cloud models for harder tasks. Local inference requires selecting a local provider; agent hosting is a separate choice.

The main limitation is memory maintenance. Stored instructions can become wrong or outdated, and a persistent AI agent still needs a model capable of retrieving, revising and cleaning its knowledge.