ConnectConnectDeepSeek Harness

MCP Client

DeepSeek Harness

Connect the DeepSeek Harness (dsh) CLI agent execution loop to XMemo using the native dsh-xmemo npm package.

Who this is for and boundary

This guide connects the DeepSeek Harness (dsh) CLI agent execution loop to XMemo persistent memory. Integration operates through the native dsh-xmemo plugin package (version 0.1.0) on npm, anchored to repository yonro/xmemo-deepseek-plugin @ e7101afb0a (https://github.com/yonro/xmemo-deepseek-plugin/blob/e7101afb0a/README.md). Important boundary: This integration is specifically for the dsh agent harness CLI tool; it is NOT an integration into DeepSeek consumer web chat (chat.deepseek.com), which does not support custom plugins or MCP extensions.

Prerequisites

Ensure the following runtime dependencies and tools are present before connecting:

  • DeepSeek Harness (dsh) CLI installed and operational.
  • Node.js runtime version 18.0.0 or higher.
  • Valid XMemo account at https://xmemo.dev.
  • Network access to https://xmemo.dev/mcp over HTTPS.

Installation

Add the native dsh-xmemo plugin to your DeepSeek Harness profile using the documented command from README lines 90-94:

dsh plugin --profile <name> add dsh-xmemo

Authorization and credential configuration

Configure credentials using environment variables or harness profile configuration. Do not place plain API keys in public project repositories:

  • Environment variable (recommended): Set XMEMO_KEY in the environment before launching dsh (export XMEMO_KEY='<your-xmemo-token>').
  • Profile configuration: Specify authentication credentials in the target profile configuration file (~/.dsh/config.json).

Load and restart behavior

Daemon restart or execution loop reload procedures are not explicitly documented in the dsh-xmemo plugin README (https://github.com/yonro/xmemo-deepseek-plugin/blob/e7101afb0a/README.md); refer to the DeepSeek Harness CLI documentation or launch tasks using your configured profile.

Connection check (read-only)

Verify connectivity with a read-only recall request before starting agent execution loops:

  • Check: Execute a read-only evaluation query through dsh without modifying memory state.
  • Success signal: dsh returns structured memory results with valid Reference and Location fields from the authorized workspace.

First useful task: Synthetic memory lifecycle

Verify memory persistence, stateless session transition, source-attributed recall, and correction:

  • 1. Save synthetic fact: Store a harness task convention ("Harness demo build target: ESM with strict typing.") to "projects/demo-dsh/conventions".
  • 2. Stateless transition: Terminate the current dsh task and start a fresh execution run using the configured profile.
  • 3. Recall with source attribution: Query "harness build target convention". Confirm returned Reference and Location.
  • 4. Correct memory: Update the remembered convention using update_memory with updated build targets.
  • 5. Optional cleanup: Soft-delete (forget) the demo convention when onboarding verification completes.

Scope of automatic capture and duplicate prevention

DeepSeek Harness execution loop memory management:

  • Loop checkpointing: dsh-xmemo automatically captures task outcomes, architectural decisions, and error resolutions at task completion boundaries.
  • Duplicate prevention: Deduplication hashing prevents re-capturing identical execution context across repeated loop iterations.
  • Scope isolation: Captured context is strictly bounded to the authenticated workspace and active repository scope.

Conflicts with existing providers and storage ownership

Runtime boundaries and storage ownership:

  • Provider conflicts: Disable default ephemeral local memory in dsh.config.json when using dsh-xmemo to avoid redundant or conflicting context injection.
  • Storage ownership: Durable memories are stored in the user-owned XMemo cloud backend; dsh stores only ephemeral execution transcripts and local run logs.
  • Offline and replay limits: Transient connectivity losses are buffered locally; extended offline periods fail write operations once the buffer is exhausted, and offline context is not shared across agents until network reconnection.

Recovery, update, and removal

Troubleshooting, update, and removal procedures for DeepSeek Harness:

  • Recovery (401 invalid_token): Verify XMEMO_KEY is exported in the shell environment running dsh.
  • Update and removal: Update and plugin removal procedures are not documented in the dsh-xmemo plugin README (https://github.com/yonro/xmemo-deepseek-plugin/blob/e7101afb0a/README.md); refer to DeepSeek Harness CLI documentation for profile management, or revoke credentials and purge stored memories at https://xmemo.dev/me.
  • Remote deletion: To purge stored memories or revoke access tokens, manage your workspace at https://xmemo.dev/me.

Source: DeepSeek Harness integration