If every OpenHands task runs in a fresh sandbox, how does memory persist?

The sandbox is isolated, but Glen lives outside it as an MCP server. The agent reads and writes durable, org-shared memory through one tool, so knowledge survives across runs.

Can different OpenHands runs and agents share what they learn?

Yes. Glen is org-scoped, so every run, every developer, and every other MCP client reads and writes the same memory.

Shared memory for OpenHands

OpenHands is an open platform for software-development agents that can edit code, run commands, browse the web, and complete end-to-end engineering tasks in a sandboxed workspace. Each task runs in a fresh environment, which is great for isolation and reproducibility but means every run starts cold: the agent relearns your project's conventions, rediscovers the same build quirks, and forgets what a previous run already figured out. Glen, shared memory for AI agents, gives OpenHands long-term shared memory as a single MCP tool, so knowledge survives across runs and is shared with every other agent in your organization.

Add Glen to OpenHands over MCP and an agent can call one tool that retrieves relevant long-term context and records new facts in a single round trip. At the start of a task it pulls what your organization already knows, the setup steps that work, the conventions to follow, the failures to avoid, so a fresh sandbox does not mean a fresh start. As the agent works, it writes back what it learned so the next run, on the same repo or a related one, begins informed rather than from scratch.

This is exactly the gap a per-run sandbox leaves. Isolation resets the environment; Glen's memory is durable and org-scoped, persisting across every OpenHands run and shared with every developer and other agent in your organization. A lesson one run learned is available to the next, and because Glen is a standard MCP server, the same memory is readable from Claude Code, Cursor, or any other MCP client, so the knowledge OpenHands accumulates feeds your whole toolchain. You connect once over OAuth or an API key and turn isolated runs into a compounding knowledge base.

FAQ

If every OpenHands task runs in a fresh sandbox, how does memory persist?: The sandbox is isolated, but Glen lives outside it as an MCP server. The agent reads and writes durable, org-shared memory through one tool, so knowledge survives across runs.
Can different OpenHands runs and agents share what they learn?: Yes. Glen is org-scoped, so every run, every developer, and every other MCP client reads and writes the same memory.