Back to catalog
Letta
Apache-2.0Letta is an open-source platform for stateful AI agents, using tiered virtual memory and PostgreSQL storage to curb recurring token inference costs.
Open-source alternative to — compare it against:
Overview
Letta provides an open-source runtime for stateful AI agents that maintain persistent memory across independent sessions. Its tiered memory hierarchy treats language models like an operating system, paging knowledge in and out of prompt windows through autonomous self-editing functions.
Why Letta instead of LangChain
LangChain provides broad orchestration primitives, document loaders, and graph pipelines, but persisting agent memory across user sessions requires assembling external checkpointers, vector stores, and custom glue code. When conversational histories expand, standard LangChain chains re-send accumulated message logs on every turn, causing context windows to balloon and compounding API inference bills. Letta replaces manual memory wiring with an integrated agent runtime that manages conversational state through database-backed tiers. Rather than flooding prompts with unpruned histories, Letta keeps a bounded working memory block in context and retrieves archival records on demand, sharply reducing token consumption.
The operational trade-off centers on architectural scope versus memory specialization. LangChain handles arbitrary multi-step graphs and broad ecosystem integrations as a general-purpose toolchain. Letta focuses specifically on stateful agents that autonomously update their persona and knowledge blocks over time. Operating Letta requires hosting an API server with PostgreSQL and pgvector, freeing teams from commercial platforms that meter runs per seat or trace. For DevOps engineers, self-hosting Letta delivers predictable compute costs while keeping agent memory on internal networks.
What it does well
- Hierarchical memory virtualization: partitions state into core memory, archival records, and message logs, preventing context exhaustion during long interactions.
- Autonomous memory editing: lets agents inspect and rewrite their memory blocks through built-in tool calls without bespoke prompt engineering.
- Model-agnostic routing: connects to cloud APIs and local engines like Ollama or vLLM, letting teams route memory updates to economical models.
- Database persistence: stores agent configurations, message histories, and embeddings in PostgreSQL with pgvector for straightforward backups.
- Multi-surface access: serves stateful agents over a REST API, TypeScript SDK, CLI, and messaging channels like Slack from one backend.
