MemoryRouterMemoryRouter
All posts

Give Claude Cowork a Shared Memory You Control

John Rood··5 min read

Here is a moment everyone who uses Claude for work knows.

On Monday, you spend twenty minutes walking Claude through a project: the client, timeline, budget constraint, and the thing your boss cares about most. The conversation is great. On Thursday, a new conversation does not reliably have those details.

MemoryRouter gives Claude access to an external, OAuth-selected memory vault. In Cowork, that access is model-directed: Claude can search or save through connector tools, but the current MemoryRouter Cowork package installs no deterministic capture hook. Nothing honestly guarantees that every turn will be recalled or saved.

That boundary matters. The useful workflow is not “just talk and everything is captured.” It is:

  • Ask Claude to check MemoryRouter when prior context matters.
  • Ask Claude to remember a concise fact, decision, preference, correction, or outcome.
  • Or let Claude propose a concise durable memory and explicitly approve the exact text before it saves.
  • Confirm the tool result instead of trusting a prose promise.

What changes with guided memory

Suppose you approved a memory that the Hendricks proposal should lead with timeline rather than price. In a new Cowork conversation, enable MemoryRouter and ask: “Check my MemoryRouter memory for the Hendricks proposal, then draft the follow-up.” Claude can call search_memories, use the approved note, and show where the context came from.

That can reduce re-explaining without turning an entire conversation into permanent storage. Routine chatter, raw transcripts, secrets, and model reasoning should not be saved. Durable memory stays concise and user-approved.

Connecting it takes about a minute

MemoryRouter connects to Claude as a custom connector:

  1. Open Customize → Connectors. On an individual plan choose + → Add custom connector. Team/Enterprise Owners first add it under Organization settings → Connectors → Add → Custom → Web.
  2. Paste https://mcp.memoryrouter.ai/mcp and leave optional OAuth client credentials empty.
  3. Click Connect, sign in with Google or GitHub, and choose the MemoryRouter vault this connection may use.
  4. In each conversation, choose + → Connectors and turn MemoryRouter on.
  5. Ask Claude to use memory_status, then complete the explicit store/recall check in the full Cowork installation guide.

The custom connector is live and can be added manually. The optional MemoryRouter Skill/plugin artifacts are for custom distribution and are not publicly approved or listed in Claude's marketplace.

OAuth scopes are the authorization boundary. memories:read permits recall and status. memories:write separately permits approved saves. A read-only Skill description is not enough by itself; read-only use must be enforced by the OAuth grant or organization tool policy.

Guided curated import from ChatGPT

Desktop Cowork can also help review selected material from an official ChatGPT export without a terminal:

  1. Request and download the official ChatGPT data export.
  2. Keep the original archive as an untouched backup.
  3. Unzip it with Finder on macOS or File Explorer on Windows.
  4. In desktop Cowork, select the extracted local folder, not the ZIP itself.
  5. Connect MemoryRouter through OAuth and choose the intended vault.
  6. Tell Claude which conversations or date/topic subset to review.
  7. Have Claude propose short, self-contained memories. Review them and explicitly approve each memory, or a clearly enumerated bounded set, before store_memory runs.
  8. Verify a selected memory from a new conversation and retain the original archive.

This is a partial, curated, model-directed workflow. It is not a complete or deterministic archive import, does not prove every conversation was processed, and does not replace a dedicated deterministic importer. Cowork should never silently save the whole export.

Shared memory for a team

A team can connect authorized users to a shared vault so approved context is available across their Claude conversations. One colleague can explicitly save a client preference; another can later ask Claude to search MemoryRouter for that client.

The sharing is real, but the use is still guided. A conversation does not become team memory merely because it happened, and a later conversation should explicitly request recall when the context matters. Use a dedicated vault when strict team or project isolation is required.

Common questions

Is my data safe? Memories are stored in the OAuth-selected private vault. Review what is stored and curate individual items in the MemoryRouter dashboard. Never paste raw Memory Keys or OAuth tokens into Claude.

Do I need to install software? The connector is remote and manually added in Claude settings. An optional custom Skill/plugin can add behavior guidance, but it does not grant permissions or make tool use deterministic.

What if I stop using it? Turn MemoryRouter off for the conversation, then disconnect/remove it from Claude connector settings. Your vault remains until you delete its contents through MemoryRouter controls.

Does Cowork automatically capture every conversation? No. The current MemoryRouter Cowork package installs no deterministic capture hook. Recall and saving are model-directed, and saving requires an explicit request or approval.

Try it free

Every plan starts with a 14-day free trial, and plans start at $20/month.

Start your free trial, add the connector, and verify one explicit save plus one explicit cross-conversation recall before relying on it for real work.

All posts