Letta recommends Agent SDK for new projects as V1 is deprecated

A recent MarkTechPost tutorial illustrates an engineering assistant whose memory can be inspected after a conversation. Its example supplies a project update, reads the stored blocks, then adds a build-status note directly from application code. It also attaches shared policy and scratchpad blocks before asking a second agent to review a proposal. These are operations shown in tutorial code, not a production result verified by CurioFuse.

The official SDK overview provides an important version boundary: V1 is being deprecated in favor of Agent SDK. V1 lists both Python and TypeScript support, while Agent SDK currently lists TypeScript only and can connect to hosted or self-hosted deployments.

For the legacy API, memory blocks are information kept in the model’s context, meaning the material it receives when generating an answer. Each block has a label, a purpose description, its contents and a character limit. The description guides how the agent uses the block. Blocks are editable by default; setting read_only to true prevents the agent from changing that block.

The current memory guide instead describes MemFS, a memory filesystem tracked with Git version control and shared across an agent’s conversations. The /init command initializes or refreshes project memory by inspecting the repository and asking about working preferences when needed. To teach a durable lesson explicitly, use /remember: the agent chooses its location and commits the update. Then /memory lets the user inspect what was saved. If the memory hierarchy drifts or grows too large, /doctor audits placement, duplication and system-prompt token usage.