cachly vs Letta: memory for your coding assistant or an agent with its own memory
Letta, formerly MemGPT, is an agent with its own memory: you talk to a Letta agent, and it learns over time. cachly gives memory to the coding assistant you already use: it keeps the lessons from fixes and decisions in your codebase.
This page compares both on facts you can check. Each row links its source.
At a glance
| Aspect | cachly | Letta | Sources |
|---|---|---|---|
| What it is | cachlyExperience for coding agents. Your assistant stores a lesson after a fix and finds it again when the same problem returns. | LettaA stateful agent harness. Letta Code agents have memory and identity, and learn by rewriting their own memory and skills. Letta was formerly MemGPT. | Sources |
| How you use it | cachlyInside the coding assistant you already use, as a plugin or an MCP server. | LettaAs its own agent: a CLI, a desktop app, chat.letta.com, Slack, Telegram and Discord. The Letta Agent SDK builds agents into TypeScript apps. | Sources |
| What gets stored | cachlyA lesson: topic, what worked, what failed, context, files and commands. Your assistant writes it, the Stop hook records reported fixes, and git history adds more. | LettaThe agent's own memory, which the agent edits as it learns. MemFS tracks all context, including memory blocks, in git. | Sources |
| MCP | cachlyAn MCP server on your machine. The editor lists 27 tools; 124 in total through one dispatcher tool. | LettaBoth ways. A hosted MCP server at api.letta.com/mcp creates Letta agents and hands work to them. Letta agents can also call external MCP tools. | Sources |
| Claude Code and Cursor | cachlyClaude Code as a plugin. Cursor, Windsurf, VS Code with GitHub Copilot, Cline, Continue.dev and Zed over MCP. A JetBrains plugin. | LettaThey connect to the hosted Letta MCP server to delegate work to Letta agents and read their replies. | Sources |
| Models | cachlyThe model your editor already uses reads the lessons. | LettaYou choose: Anthropic, OpenAI, Google Gemini, Mistral, DeepSeek, Groq and more, with your own API keys. | Sources |
| Where data lives | cachlyHosted by cachly: one instance per brain at Hetzner in Nuremberg, Germany. By default, cachly computes embeddings in the EU. The API runs behind Cloudflare, a US company. | LettaLetta Cloud is the default and keeps agents' memory, identity and conversations. The local runtime keeps all agent state on your machine. | Sources |
| Run it yourself | cachlyHosted service. The MCP server and the hooks run on your machine and talk to your instance. | LettaYes: run agents locally, or host the Letta App Server on your own machine. No Letta account required. | Sources |
| Teams | cachlyTeam plan: €49 a month for 5 seats. A lesson one developer stores reaches the whole team. | LettaShared memory: Git repositories in Letta Cloud that several agents read and maintain. Members of an organization need Teams Pro. | Sources |
| License | cachlyMCP server and hooks: Apache-2.0. The hosted service is closed source. | LettaLetta Code: Apache-2.0. Letta Cloud is a managed service. | Sources |
| Pricing | cachlyFree €0: one brain, 25 MB, 500 recalls a month. Dev €19, Pro €49, Business €199 a month. Team: €49 a month for 5 seats. | LettaFree with your own API keys, up to 3 stateful agents. Pro $20 a month. Teams Pro $20 per seat a month. Developer plans for the API are usage-based. | Sources |
| Community | cachlycachly-mcp: public repository since April 2026. | LettaMemGPT, a research project from UC Berkeley, became Letta in September 2024. Letta Code has been public on GitHub since October 2025. | Sources |
As of . Every statement about Letta comes from the linked source, read on that day. Products change; if you spot something outdated, write to [email protected].
Two different jobs
With Letta, the agent is the product. Letta Code runs in a terminal, a desktop app or a chat, and its memory belongs to that agent. Claude Code or Cursor can hand work to a Letta agent over MCP.
With cachly, your assistant stays the same. Claude Code, Cursor or Copilot keep working as before, and cachly adds what the last session learned to the next one.
When to pick Letta
- You want an agent with its own memory and identity that learns over long stretches of time.
- You want to reach the same agent from a terminal, a desktop app, a browser or Slack.
- You want its memory as files in git, in your own repository.
- You want to choose the model yourself and run everything on your own machine, open source under Apache-2.0.
When to pick cachly
- You keep your coding assistant and want it to remember fixes, dead ends and the files involved.
- You use Claude Code and want matching lessons next to each prompt, through a hook.
- Your data should be stored in Germany, with embeddings computed in the EU.
- A flat monthly price fits you: free with 500 recalls a month, or from €19 a month.
- You want to try it before you create an account.
Try cachly
In Claude Code, cachly comes as a plugin with three hooks. Try it without an account: the first tool call creates a test brain on EU servers. The commands are in Hooks that carry experience.
For Cursor, Windsurf, VS Code and other editors, MCP integration shows the one-command setup. What leaves your machine, and where it lands: EU hosting and your data.
Sources
Other comparisons: cachly vs Supermemory · cachly vs Zep · cachly vs claude-mem · cachly vs Basic Memory · cachly vs mem0 · All docs