# Cross-Agent Memory > Source: > Published: 2026-08-28 13:38:38+00:00 Agents forget. Worse, they forget separately. You explain your stack to Claude Code on Monday, explain it again to Codex on Tuesday, and explain it a third time to whatever you open on your phone. Each tool has its own notion of what it knows about you, and none of them are the notion you actually maintain. Lanes Link treats memory as yours rather than the agent's. It is served from the endpoint you run, so what one session writes is served back to every later session, including a session in a different agent entirely. The same is true of skills: your reusable procedures, exposed as MCP prompts, available wherever you are working. ## How Lanes Link does it **Memory is a first-class layer, not a file convention.** It is accumulated knowledge about you and your work, worth searching before an agent concludes it knows nothing. The search is a tool call, so any MCP client can make it.**Writing is a separate grant.** Reading memory and adding to it are different capabilities. An agent that should draw on your context does not automatically get to edit it.**Skills are yours, not the harness's.** Reusable procedures live at the endpoint and are exposed as MCP prompts, so a procedure you refine once shows up in every client rather than being re-pasted per tool.**The vault covers the sensitive half.** Passwords and API keys are handed out one item at a time, and nothing lists what else is in there. Knowing one secret does not reveal the shape of the rest.**No credential required.** Memory, skills, and the vault are the owner layer. They serve without any account being connected, so an endpoint with zero integrations is already useful on day one. ## The workflow Ask an agent to remember something and it writes to the endpoint. The next session, in any client you have registered, searches the same store and finds it. There is nothing to sync and no per-tool copy to keep current, because there was only ever one copy. This changes how the first minute of a session goes. Instead of re-establishing context, the agent looks it up. A convention that lives in memory, a preference you stated once, a decision you made three weeks ago in a different tool: all of it is reachable through the same search, and all of it stays behind your permission boundary and your audit log. ## Set it up Install it, start the endpoint locally, then register it everywhere you work: ``` bash $ bun install -g @lanes-sh/link $ lanes link profile add personal --default $ lanes link start ok serving http://127.0.0.1:7337/mcp profiles: personal ``` Then, in another shell: **Run** It registers`lanes link mcp add` .`lanes-link` with Claude Code and Codex, and installs a skill plus a scout agent so the agent knows to search your memory before assuming it knows nothing.**Write something worth keeping.** Ask an agent to remember a convention, a preference, or a decision.**Open a different client and ask for it.** Same endpoint, same store, same answer. ## Further reading [Lanes Link: one endpoint for everything an agent needs](/blog/introducing-lanes-link): what the endpoint serves, and where state actually lives.[Connect your accounts once, for every agent](/use-cases/one-endpoint-for-every-agent): the connection layer that sits beside memory.[Give agents scoped access, with an audit trail](/use-cases/scoped-account-access-for-agents): how writes are granted, and how every call is recorded.[Keep work and personal accounts apart](/use-cases/separate-work-and-personal-agent-access): profiles share no state, so work memory stays out of personal sessions. Browse the source at [lanes-sh/link](https://github.com/lanes-sh/link), Apache-2.0, or read the [announcement](/blog/introducing-lanes-link).