A knowledge base in files you own versus a research-driven agent framework. The right choice depends on whether you want memory for your AI tools or infrastructure for agents you are building.
Basic Memory gives your existing AI tools. Claude, ChatGPT, Cursor, Codex. A persistent, shared knowledge base stored as plain Markdown you can read and edit. Letta (formerly MemGPT) is a full agent framework: you build stateful agents inside it, and it manages their memory, tools, and execution. If you want your AI tools to remember your work, use Basic Memory. If you are engineering autonomous agents and want managed state and a runtime, Letta is serious infrastructure for that.
| Basic Memory | Letta | |
|---|---|---|
| What it is | Knowledge base for you and your AI tools | Agent framework with persistent state |
| You can read what is stored | Yes. Plain Markdown files | Partially. Through the framework |
| Works with your existing AI tools | Yes. Any MCP client (Claude, ChatGPT, Cursor…) | No. You build agents inside Letta |
| Open source | Yes (AGPL-3.0) | Yes (Apache 2.0) |
| Origin | Built for human + AI knowledge work | MemGPT research project (UC Berkeley) |
| Funding model | Bootstrapped | $10M VC |
| Pricing | Free local; Cloud from $15/seat/month | Open source; Letta Cloud usage pricing |
Letta grew out of the MemGPT paper, which treated an LLM’s context window as virtual memory. Genuinely innovative research, and the team continues to think deeply about agent memory architecture. Today Letta is a framework: agents, tools, state management, and a runtime, with memory as one managed subsystem inside it.
Basic Memory sits at a different layer. It is not a framework you build inside; it is a knowledge base that plugs into the tools you already use via the Model Context Protocol. Your notes, decisions, and research live in Markdown files on your machine (or your private cloud), and every MCP-compatible assistant reads and writes the same graph.
The practical test: if you are writing code that instantiates agents, Letta is a candidate. If you are a person (or team) who wants Claude and Cursor to stop forgetting your project context, that is Basic Memory’s job. No framework required.
Even as Letta moves toward more file-based approaches, agent memory remains several layers away from "open a folder and read it." Basic Memory starts there: transparency is the architecture, not a feature. Notes are Markdown with semantic links; the knowledge graph is derived from text you can read, edit, and version with git.
For agent builders this cuts both ways. Letta’s managed abstractions do work for you that Basic Memory leaves in your hands. For knowledge ownership, files win: nothing about your memory is locked inside a runtime.
Open source to run locally. Cloud from $15/seat/month with a 7-day trial.
Knowledge tools built to last
© 2026 Basic Memory - Basic Machines