Deep agent-memory design
Letta comes from the MemGPT line of research and describes agents whose memory, identity and capabilities evolve with experience.
Letta is a platform for building stateful agents whose memory is part of the agent, plus the Letta Code coding agent. MeshTale is one governed memory shared by the agents you already use. A different job, not a weaker one.
Last checked 29 Oct 2026, from Letta's own pages.
Letta 0MeshTale 0Different jobs, both valid.
A platform for stateful agents whose memory lives inside the agent, plus Letta Code. These are the things it does well, taken from the sources listed at the bottom of this page.
Letta comes from the MemGPT line of research and describes agents whose memory, identity and capabilities evolve with experience.
Memory blocks sit in the agent's context window and agents update them with memory tools. Blocks can be shared across agents.
In Letta Code, memory lives in a git-backed filesystem and every memory edit is committed.
Apache-2.0 and self-hostable, with a free plan and a low-cost Pro plan.
"Not stated" means we checked Letta's pages and found nothing either way. It does not mean the feature is missing.
| What you are comparing | Letta | MeshTale |
|---|---|---|
| What it is | A platform for stateful agents with built-in memory: an agent runtime, API, a visual development environment and the Letta Code coding agent. | One governed memory for every person and AI agent on a team: hosted app, MCP, REST API with workspace-scoped keys, and a command-line sync tool. |
| Who it is for | Developers building stateful agents, and individual developers using Letta Code. | Teams and agencies that share context across people and AI agents, usually for several clients. |
| Open source and self-hosting | YesApache-2.0; self-host with the Letta server or Docker. Self-hosted agents are not reachable from Letta's hosted chat. | PartlySource-available under FSL-1.1-MIT. Runs as a hosted service; a supported self-host offering is not part of the plans. |
| Separate spaces per client | PartlyPer agent, with memory blocks shareable across agents; access control on Teams Pro. Per-client separation is not stated. | YesOne workspace per client, separate by default: one client's context never reaches another client's agent. |
| Review before recall | Not statedNot stated for memories. Human approval is documented for agent tool calls. | YesNew memories can wait in a review queue for a person's OK; trust rules let matching items skip it. |
| Personal-data masking | Not statedNothing either way on the pages we checked. | YesEmails, phone numbers and ID numbers are detected and can be masked before storage, with per-workspace data policies. |
| When memories disagree | PartlyAgents edit their own memory; Letta Code keeps every memory edit as a git commit. | YesConflict alerts. When two memories disagree a person picks which is accurate (or keeps both with context); the other is marked superseded and the decision is logged. |
| Connectors to business tools | Not statedNothing either way on the pages we checked. | Yes26 connectors, including Slack, Microsoft Teams, Gmail, Notion, Google Drive, Jira, GitHub, HubSpot, Granola and Zoom, plus webhooks and export. |
| MCP, works with AI tools | YesAs a client: Letta agents can use tools from remote MCP servers. | YesTwo tools to connect; works in Claude Code, Cursor, Codex, Gemini, Grok, Windsurf, Zed and anything that speaks the protocol. |
| Built-in chat with citations | PartlyFor developers and individuals: Letta's hosted chat and the Letta Code app. Citations are not stated. | YesChat with citations on every plan; every answer points back to its sources. |
| Audit log | Not statedNothing either way on the pages we checked. | YesEvery recall is logged: which agent, which workspace, what came back, when. History kept 7 days to 2 years by plan. |
| Compliance claims | Not statedNothing either way on the pages we checked. | None claimedMeshTale makes no certification claims. |
| Pricing model | Free plan; Pro and per-seat Teams Pro subscriptions; an API plan with usage-based charges; Enterprise is custom. | Free tier, no card required. Paid plans: Pro, Agency, Scale and Enterprise. |
Letta compliance claims are quoted as Letta states them. MeshTale facts come from MeshTale's own product and pricing pages.
If the left column sounds like you, Letta is probably the better fit, and that is fine.
Worth knowing. Letta is an agent runtime more than a shared team memory, so the two are often not either-or.
Every Letta fact on this page comes from Letta's own website, docs or official GitHub repository, opened on 29 October 2026. Products change; if something here is out of date, tell us and we will fix it.
Building it yourself, or keeping memory on your own computer? See those comparisons.
Mostly an agent platform: memory is built into the agents you create with it, and Letta Code is its own coding agent. MeshTale is a memory that other agents connect to.
Letta agents can use tools from remote MCP servers, and MeshTale works with anything that speaks the protocol. We have not published a Letta guide yet.
Letta is open source under Apache-2.0, and its hosted service has a free plan. MeshTale also has a free tier, no card required.
One governed memory, a workspace per client, and a log of every answer. Start on the free tier and see if it fits.
Free tier, no card required.