Developer shares Claude Code workflow with local MCP shared memory

On October 18, 2026, a developer published a write-up on Juejin describing a workflow built on Claude Code and Utility.app for connecting to LLM APIs. The central idea is to use a local MCP service as a shared memory store, allowing multiple Agents to read and write the same memory across sessions and thereby stay consistent in behavior. Because the MCP protocol is designed to be pluggable, the memory store can be attached to any Agent that supports MCP.

The developer organizes memory into hot, warm, and note layers: the hot layer keeps recent memories, the cold layer retains what remains useful over the long term, and content ultimately settles into a note layer for permanent storage and human reading. The local MCP service ships with two built-in tools, remember and recall, which handle writing and retrieving memories respectively; Agents operate on memory by calling these tools.

To address long-context contamination, he proposes injecting the core prompt into every latest user message, countering the tendency of system prompts, rules, and tool descriptions to be diluted by later content in long conversations. He considers opening a new session each time to be cumbersome. In addition, he places a proxy gateway between the Agent and the upstream LLM provider, making the black-box system transparent while logging requests and quantifying Token costs. He also notes that compressing historical messages breaks the cache. The report does not include a specific version number or quantitative figures for the workflow.