# Durable Thinking: Keep a reasoning record across sessions

URL: https://research-memex.org/docs/toolkit/durable-thinking
Description: Store explicit reasoning notes, branches, and revisions in your own Cloudflare deployment, then retrieve them across MCP sessions.



*Keep a written reasoning record after the session ends.*

<ToolMeta slug="durable-thinking" runtime="Cloudflare Workers and SQLite-backed Durable Objects" />

Durable Thinking stores the steps a model submits to an MCP tool. You can retrieve a sequence later, follow its branches and revisions, or delete it.

It records tool-visible text, not a model's hidden reasoning. A recorded explanation still needs evidence and human review.

The project adapts the MCP reference Sequential Thinking server. The [upstream license and attribution](https://github.com/linxule/durable-thinking#license) explain that relationship.

## Choose the right kind of record [#choose-the-right-kind-of-record]

| Tool                                                         | What it keeps                                            | When it fits                                                     |
| ------------------------------------------------------------ | -------------------------------------------------------- | ---------------------------------------------------------------- |
| [Sequential Thinking](/docs/toolkit/sequential-thinking-mcp) | Explicit steps in the server process's memory            | You need a revisable record during the current run               |
| Durable Thinking                                             | Explicit steps in a persistent Cloudflare store          | You need to revisit a sequence after a session or process ends   |
| [Memex](/docs/toolkit/memex-plugin)                          | Searchable project memos and saved collaboration context | You need decisions and context across a broader research project |

Durable Thinking is a separate deployment, not an automatic upgrade to Sequential Thinking or a Memex plugin.

## Use it in a research task [#use-it-in-a-research-task]

Ask the model to write out competing explanations for a finding, marking revisions and alternative branches. Save the returned sequence ID with your project notes. In a later session, use that ID to retrieve the record and check it against the evidence.

The server exposes `sequentialthinking`, `get_thought_history`, and `delete_thought_sequence`. There is no sequence-listing tool, so keep the ID if you want to return to the work.

## Deploy and connect [#deploy-and-connect]

Deploy your own copy using the [upstream deployment guide](https://github.com/linxule/durable-thinking#deploy-your-own). You need a Cloudflare account, a GitHub OAuth app, and the runtime secrets described there. This is a private, single-user deployment, not a public hosted service or a local npm server.

Connect a compatible remote MCP client to your deployment. The [client connection guide](https://github.com/linxule/durable-thinking#connect-your-clients) distinguishes browser sign-in from bearer-token access and explains which endpoint to use.

Clients supporting MCP Apps can show a card for each step and the completed sequence. Other compatible clients receive text results. Rendering support and authentication support depend on the client.

## Protect the record [#protect-the-record]

Submitted text is stored in your Cloudflare deployment, not only on your computer. Avoid including credentials or unnecessary personal data. Authentication controls entry to the server; a sequence ID controls access to that sequence within it. Keep both credentials and sequence IDs private.

Sequences remain stored until explicitly deleted unless you configure expiration. Deleting a sequence requires confirmation and permanently removes its stored text. See the [retention and privacy guidance](https://github.com/linxule/durable-thinking#retention-and-privacy) before using sensitive research material.