Introduction

DeepSeek Harness (DSH) uses a pluggable architecture. When building multi-turn conversation or long-term memory scenarios, preserving information after a session ends is a common pain point. dsh-auto-memory is a session-level memory plugin designed to solve the problem of losing cross-session decisions and conclusions. Through automatic wrap-up reminders and memory index maintenance, it helps the model remember what truly matters.

Core Features

The plugin mainly includes the following capabilities:

  1. Wrap-up Reminder: At the start of a session, the project memory index is injected into the model. When the number of index lines or the file size exceeds the configured soft limit, it proactively reminds the user to organize it.
  2. Wrap-up Marker: If a session is destroyed without completing wrap-up, a marker is written to the state directory. At the start of the next session, it reminds the user to complete wrap-up; if wrap-up has already been performed, the marker is cleared.
  3. Memory Index: Maintains a MEMORY.md file in the project root directory, with one topic per line, while the memory/ directory holds detailed content.
  4. Scaffolding: Automatically initializes the MEMORY.md index and the memory/_TEMPLATE.md template at session start, and installs a user-level wrap-up skill into ~/.dsh/skills/auto-memory.
  5. Fully graphical configuration: All configuration is completed through switches, dropdowns, or text fields in the settings panel, without manually modifying YAML files.

Installation and Enabling

  1. Ensure the environment meets the requirements: DSH (DeepSeek Harness) and Node.js 20+.
  2. Run the installation command:
    dsh plugin --profile web add <仓库地址或本地目录>
  1. Restart DSH after installation is complete.
  2. Go to “Settings → Plugins”, find AutoMemory, and configure it.

Typical Usage

After installing and enabling the plugin, the panel provides configuration items for the memory index file name (default: MEMORY.md) and the soft limit (default: 150 lines and 20 KB). At session start, the plugin automatically injects the MEMORY.md index into the model and issues an organizing reminder when the content exceeds the limit. At the same time, it automatically initializes the project scaffolding files and installs the wrap-up skill to guide the model in performing wrap-up.

Notes

  • Module Changes: The original remember submodule (including conversation extraction, LLM summarization, NDC, and consolidation) has been completely removed. The reason is the stability issue of the summarization model with SKIP decisions and the relatively high cost of fixing it. This plugin now only retains the autoMemory core features, which have no model dependency.
  • Environment Dependencies: Requires a Node.js 20+ environment. The DSH runtime provides peer dependencies (@deepseek-ai/cordis, @deepseek-ai/schemastery, @deepseek-ai/dsh-settings).
  • Panel Display Issues: If the panel does not appear, try a hard browser refresh (Ctrl+Shift+R) and confirm that DSH was restarted after installation.

Conclusion

The plugin lowers the barrier for DSH users to maintain session memory through graphical configuration and automated scaffolding. The related source code and directory can be viewed via the link below.