Letta

Key features

• Stateful memory blocks Editable memory sections persist per agent across sessions, so instructions, user facts, and project context survive after the chat ends. The agent itself can update them as it learns, which keeps stored context current without manual edits.[1]

• Archival recall Long-term memory stores searchable history the agent can retrieve when current context runs short, keeping long engagements coherent. This is the layer that prevents month-long deployments from degrading into amnesia.[1]

• Tool use with approval flows Agents call custom tools and integrations, with documented approval steps that keep consequential actions under human control. Teams decide which calls run freely and which wait for a person.[1]

• Files as context Uploaded and referenced files become working material the agent can read during reasoning, grounding answers in project documents. Document-heavy workflows gain answers tied to actual sources rather than model guesses.[1]

• Multi-agent groups Groups coordinate several agents on shared work, with jobs scheduling background execution outside the interactive loop. Long-running assignments progress without a user watching the chat.[1][2]

• Voice agent support Documented voice models and agent types extend the same memory-backed loop to spoken interaction use cases. Spoken assistants keep the same persistent memory as their text counterparts.[1]

• Model flexibility with tracing Teams can attach capable or reasoning models from major providers and inspect behavior through provider tracing. Switching models does not mean rebuilding the memory layer around them.[1][2]

• MCP and SDK surfaces Model Context Protocol support plus TypeScript and Python SDKs connect agents to existing codebases and tooling. Existing systems expose capabilities to agents through a documented standard instead of bespoke glue.[1]

Who will love it

• Support copilots, research aides, and personal productivity agents gain continuity when memory blocks and archival recall persist across sessions. If user-specific context is the product's value, this is the layer that holds it.[1][2]

• One server, one API, and two SDKs replace scattered session stores, cron scripts, and per-app tool wiring. Platform teams get a single surface to secure, observe, and scale.[1]

• The documented Docker and npm local setup against Postgres gives full data control to teams that already run their own infrastructure. Regulated or privacy-sensitive deployments keep memory stores inside their own boundary.[1]

Who should avoid it

• The documented path is APIs, SDKs, and server setup; no verified no-code builder exists, so visual-first teams will stall. Every customization assumes someone on the team writes code.[1]

• The cloud tiers were confirmed but exact per-plan prices were not verifiable, so fixed-budget procurement cannot close on this review alone. Check the live pricing page before committing.[1]

• If the task fits one stateless call, running a stateful agent server adds moving parts without a matching payoff. A direct model API call stays cheaper and simpler there.[1]

Compatibility and requirements