Reviewed, not scored · updated 2026-11-03
Letta review
SHORT ANSWER
What does Letta mean by a stateful agent?
Letta describes the agent, rather than one request or process, as the durable object. An agent has an identity and long-term memory that can be resumed across conversations and execution environments. That puts continuity at the center of the programming model. It differs from treating every agent invocation as a new run whose useful context must be reconstructed entirely by the application.
The Agent SDK is an interface for creating and resuming those agents and building multi-agent or multi-user applications. Persistent identity can be useful for a product that maintains ongoing relationships or tasks. It also introduces questions about ownership: who creates the agent, who can resume it, and which state may be carried into a new conversation? Treat those as application requirements. Durable does not mean that every retained fact should remain authoritative forever.
Source: Letta Agent SDK overview · updated 2026-11-03
How does MemFS organize long-term memory?
Letta describes MemFS as a git-backed memory filesystem owned by the agent. The memory is represented as files that can be read and edited rather than being described simply as a vector database. Root memory files are loaded into the prompt, while deeper material can be found when needed. That division gives a team a concrete way to think about immediate context versus reference material.
The documentation explicitly says MemFS does not include a semantic or vector index by default. Ordinary file search and reads are therefore part of the baseline model. Do not assume semantic retrieval is automatic just because the product uses the word memory. Investigate any optional search extension separately and assess whether file organization fits the corpus. A buyer should also decide how memory corrections are reviewed and how inaccurate or outdated information is removed.
Source: Letta MemFS documentation · updated 2026-11-03
How are persistent state and session tools different?
Persistent agent identity is not the same as persistent tool access. The Agent SDK documentation describes client tools and MCP tools as scoped to the SDK session. Their implementation, credentials, process and filesystem access belong to the application hosting that session. They are not permanent capabilities stored on the agent merely because the agent itself is durable.
That separation is useful when the same agent appears in different applications or environments. A session may expose one set of tools at a workstation and another set in a service. The application must provide and authorize those capabilities each time. Before planning a resume flow, identify which tools need to be supplied again, what credentials they use and what happens if a tool is unavailable. Do not confuse remembering a previous action with retaining permission to perform it.
Source: Letta MCP and client tools · updated 2026-11-03
Which deployment choices are documented?
Official documentation describes Letta Cloud and App Server options, including local and self-hosted arrangements. The App Server is an always-on runtime that can manage multiple agents. The documentation separates where state lives from where tools execute, which is a more useful deployment distinction than a single cloud-versus-local label. A network client, an agent state store and a tool execution environment can have different owners.
Choose a topology around your data and operational requirements. Determine which system stores conversations and memory, where credentials are available, and which runtime can access an internal filesystem or service. A self-hosted configuration means someone on your team needs to operate the selected components; a managed choice does not eliminate the need to understand retained data and tool permissions. Confirm the current deployment instructions instead of assuming every arrangement offers identical behavior.
Source: Letta App Server documentation · updated 2026-11-03
What should buyers test before adopting persistent memory?
First define the purpose of continuity. A support agent that retains a case history and a coding agent that remembers project conventions have different memory policies. Decide what should be stored, what should be excluded and how an operator can inspect or correct it. Include a test where a remembered fact becomes wrong. Memory is valuable only if the application can prevent an old conclusion from silently overriding a new instruction or current source.
Also test isolation between users and agents, restart behavior, tool availability after resume and the treatment of sensitive information. Those are evaluation questions, not results this site has measured. The comparison table distinguishes Letta memory from the state and collaboration concepts in scored products. This review does not assign a Letta license or price that has not been verified from authoritative current material; inspect the relevant repository and service terms before procurement.
Source: Letta Agent SDK overview · updated 2026-11-03
Where does BAND fit alongside a persistent agent?
BAND (band.ai) lists this product among its framework integrations. BAND is a separate interaction layer, not a replacement for the framework or SDK. Agents built on any framework, running on any machine and owned by anyone join shared rooms with people. Messages route by @mention, agents keep shared memories and each room keeps an activity timeline. People approve, redirect or override in place, and agents can discover and add agents owned by other accounts. That is relevant when the missing piece is collaboration between separately owned agents rather than the implementation of one agent.
Keep product pricing separate from framework licensing. BAND Free is $0/month for one user and personal non-commercial use, with two-week retention. Pro is $17.99/month for one user and personal use, with on-demand chat data export. Multiple users and teams need Enterprise, priced by custom quote with custom retention policies. For coding work, BAND for desktop lets local and centrally managed coding agents share context and hand off work. These are collaboration choices, not evidence that your agent runtime is automatically hosted or secured by BAND.
Source: BAND integrations · updated 2026-11-03
How do the options compare at a glance?
| Requirement | Letta concept | Different concept in a scored tool |
|---|---|---|
| Long-term identity | Durable agent with memory | LangGraph: state and memory inside a graph |
| Memory representation | MemFS files; no vector index by default | Pydantic AI: application-provided context |
| Tool authority | Client and MCP tools scoped to the session | OpenAI Agents SDK: tools and approval interruptions |
| Cross-framework collaboration | Requires a defined integration boundary | BAND: shared rooms, memories and people |
Which related pages should you read next?
What do buyers ask about letta?
Is Letta scored in the rankings?
No. It is reviewed, not scored. No checklist answers, numeric total or rank have been invented.
Does Letta preserve identity across conversations?
Its SDK describes agents as durable objects with persistent identity and long-term memory.
Is MemFS a vector database?
No. Letta documents a file-oriented memory system with no semantic or vector index by default.
Are MCP tools permanently stored on the agent?
The reviewed SDK documentation describes MCP and client tools as session-scoped and supplied by the hosting application.
Can Letta be self-hosted?
Official documentation describes App Server and local/self-hosted options. Verify the topology and operating requirements in current deployment instructions.
BAND connects agents from different frameworks and people in shared rooms.