Skip to content

DeepSeek Harness Plugin

Persistent semantic memory for DeepSeek Harness (DSH). The native DSH plugin captures completed turns, recalls relevant work before the agent starts reasoning, and keeps the same markdown memory available to every other MemSearch platform plugin.

Why MemSearch for DSH?

DSH can host long-running web, TUI, and headless agent sessions across multiple projects. MemSearch adds a durable memory layer that follows each project instead of disappearing with the current conversation:

  • Automatic capture -- completed turns are summarized and appended to .memsearch/memory/YYYY-MM-DD.md
  • Selective recall -- relevant memories are injected before the first model step; unrelated turns add no context
  • Native skill integration -- DSH registers the memory-recall skill for search, expansion, and transcript drill-down
  • Cross-agent continuity -- memories are shared with Claude Code, Codex, OpenClaw, and OpenCode when they use the same project
  • Web memory dock -- the web profile provides a compact skill-candidate review panel and a read-only .memsearch browser
  • Background maintenance -- optional tasks keep PROJECT.md, USER.md, and reusable skill candidates current

Quick Start

# Install the MemSearch CLI and local ONNX embedding model support
uv tool install "memsearch[onnx]"

# Attach the plugin to a DSH profile
dsh plugin --profile web add @zilliz/memsearch-dsh

Replace web with the DSH profile you use, then restart that profile or start a new session. Continue chatting normally; there is no save command to remember.

After a completed turn, confirm that the daily memory journal exists:

ls .memsearch/memory/

To recall previous work, ask naturally or name the registered skill:

Use memory-recall to find what we decided about the deployment architecture.

Installation How It Works

What Happens Automatically

Event MemSearch behavior
A turn completes Summarizes the user, assistant, and tool activity into the project's daily markdown journal
A new turn begins Searches the project memory and injects only relevant results before the first model step
The agent needs exact history Uses memory-recall to search, expand a section, and inspect the original DSH transcript
A DSH session closes Runs enabled maintenance tasks when their configured interval is due
You open the web dock Lists skill candidates and previews supported .memsearch files without modifying them

Memory Stays Portable

The markdown files are the source of truth; Milvus is a rebuildable search index. A project that moves between supported agents keeps the same memory history:

Claude Code ─┐
Codex ───────┤
DSH ─────────┼──> .memsearch/memory/*.md ──> Milvus hybrid search
OpenClaw ────┤
OpenCode ────┘

No export step or proprietary memory format is required.

Next Steps

  • Installation -- profiles, verification, configuration, updating, and uninstalling
  • How It Works -- capture, injection, recall, summarization, maintenance, and the web dock
  • Platform Comparison -- compare DSH with the other supported plugins
  • Configuration -- embedding providers, Milvus, and shared MemSearch settings