Compare/Agent platforms
ContxOS vs Letta: Memory for Your Existing Agent, or a Platform for Stateful Agents?
Letta is a platform for building stateful agents with git-backed memory files. ContxOS adds a shared project memory to the agents you already use. How they differ.
Jose R. Gonzalez, CPA · Founder of ContxOS
Updated · 2 min read
In short
Letta, formerly MemGPT, is a platform for building agents that own their memory. Memory is a git-backed repository of files, some loaded every turn and the rest on demand, with Letta Code, a cloud platform and an Agent SDK on top. ContxOS doesn’t replace your agent. It gives the agents you already use, Claude.ai and Claude Code among them, a shared project record over MCP.
Side by side
| Capability | ContxOS | Letta |
|---|---|---|
| What it is | A hosted knowledge base that AI agents read and write while they work on your projects. | A platform for building stateful agents (Letta Code, cloud platform, Agent SDK). |
| Which agent | Yours: any MCP client. | Agents built and run on Letta. |
| How memory is stored | Typed entries (decision, result, plan, TODO, bug, finding, note…) joined by typed links (supersedes, resolves, implements, supports, refutes…), each with an authority level. | Git-backed memory files. System files load every turn, others on demand. Every change commits. |
| Search | Semantic search with freshness and trust signals. | Not by default (a "MemFS Search" mod adds it). |
| Procedures | Workflows (stages, steps, per-client variants). Runs track position, verified facts and holds outside the chat, and proven steps can go to subagents. | Skills stored in memory. Agents can write their own subagent orchestration scripts (Sep 2026). |
| Session continuity | Structured handoffs (accomplishments, pending work by priority, blockers by severity, decisions). "resume" restores one in a single step. | The agent is persistent, so its memory carries forward. |
| Self-host | Hosted only. | Yes. Open source (Apache-2.0) and a self-hosted App Server. |
| Price | Free plan with one project and one seat. Paid plans add projects and seats (see contxos.com/pricing). | Free (3 agents, your own keys); Pro $20; API $20 plus usage; Teams Pro $20/seat. |
Competitor details come from their own documentation as of September 29, 2026 (see Sources). Products change quickly; tell us if something here is out of date.
Different starting points
Letta asks you to run your agents on Letta, and in return the agent itself is stateful. ContxOS assumes you already have an agent you like, whether that is Claude.ai, Claude Code or another MCP client, and gives it a memory it did not have. If you are designing a new agent product, Letta is a platform to build it on. If you want your current agents to stop forgetting, ContxOS is the shorter path.
Which one fits
ContxOS fits when
- You want to keep using Claude.ai or Claude Code as they are.
- You want one memory shared across several clients and teammates.
Letta fits when
- You are building your own stateful agents and want memory, runtime and SDK in one platform.
- You want everything open source and self-hostable.
Sources
- Letta documentation — Letta
- MemFS — Letta
- Context repositories — Letta, February 12, 2026
- Introducing the Letta Agent SDK — Letta, August 17, 2026
- Dynamic workflows — Letta, September 28, 2026
- Letta Code pricing — Letta
Frequently asked questions
Can a Letta agent use ContxOS?
Letta agents can call MCP tools, so in principle yes. We have not tested it.
Give your agent a memory that lasts
Connect ContxOS to Claude in about a minute. The free plan includes one project.
Product names belong to their owners. ContxOS is not affiliated with the companies named here.