SynaBun is an open-source, self-hosted AI workspace and persistent memory MCP server for the coding agents you already use. Letta, formerly MemGPT, is an agent platform built around Letta Code, whose agents keep their memory as Markdown files in git.
Last updated
| Feature | SynaBun | Letta (MemGPT) |
|---|---|---|
| What it is | AI workspace and memory MCP server for Claude Code, Codex, OpenCode and other MCP clients | Agent platform: Letta Code agents, the Agent SDK and Letta Cloud |
| Memory model | Memories with categories, project labels and importance; recall combines keyword search and embeddings | MemFS: Markdown memory files in git; memory blocks stay in the prompt; no vector index by default |
| MCP | MCP server with 158 tools (141 on by default) | Agents use MCP servers as clients; Letta Cloud has a hosted MCP server with 6 tools |
| Default embeddings | all-MiniLM-L6-v2, computed on your machine | None for memory files; message search is full-text locally, full-text, vector or hybrid on Letta Cloud |
| Where memory lives | SQLite on your machine | Files under ~/.letta on your machine, or Letta Cloud |
| SDK | None needed: it plugs into the agents you already use | Agent SDK in TypeScript; the legacy V1 SDK in Python and TypeScript |
| Managed cloud | No | Letta Cloud: free plan, Pro from $20 a month |
| License | Apache 2.0 | Apache 2.0 |
| Origin | First commit in February 2026 | The MemGPT research paper (UC Berkeley, October 2023) |
Letta facts checked on October 21, 2026 against its repositories, documentation and pricing (sources at the end). SynaBun facts are from its source code.
A full agent platform. Letta gives you the agent itself: Letta Code agents, an Agent SDK to build your own, and a hosted Letta Cloud to run them. If you are building an AI product with agents of its own, Letta is the framework for it; SynaBun has no SDK and does not run agents for other people's products.
Memory the agent edits itself. In MemFS an agent's memory is a set of Markdown files in git: the files under system/ are in the prompt on every turn, and the agent reads and edits the others when it needs them. That makes its memory easy to read, diff and version.
A hosted option. Letta Cloud runs agents for you, with a free plan and paid plans. SynaBun runs only on your machine.
Research lineage. Letta grew out of MemGPT, the 2023 UC Berkeley paper "MemGPT: Towards LLMs as Operating Systems".
Your agents, not new ones. SynaBun plugs into the coding agents you already use. It configures itself for Claude Code, Codex, Gemini CLI and OpenCode, ships side panels for Claude Code, Codex and OpenCode, and any other MCP client can connect to its server.
Search by meaning, locally. SynaBun embeds every memory with all-MiniLM-L6-v2 on your machine and recalls with keyword search and embeddings together, with no API key. Letta's memory files have no vector index by default.
More than memory. The same install adds the Assistant that dispatches Claude Code, Codex and OpenCode worker agents, browser automation (46 Playwright tools), social media tools for seven platforms, a whiteboard, autonomous loops and Discord tools: 158 MCP tools in all.
Setup. One command (npm install -g synabun, then synabun) or the desktop app, with SQLite and no server to run.
Pick Letta if: you are building a product with its own agents, you want an agent SDK and a hosted runtime, or you want memory as versioned Markdown files the agent maintains.
Pick SynaBun if: you work with Claude Code, Codex, OpenCode or Gemini CLI every day and want them to share one memory that stays on your machine, plus the Assistant, a browser and automation in the same install.
Yes, and they do not overlap much: Letta can run the agents of a product you build, while SynaBun gives your own coding agents memory and tools as you build it. Letta agents can also connect to MCP servers, SynaBun's included.
Checked on October 21, 2026.
One command. SQLite + local embeddings. 158 MCP tools.
Read the docs GitHub