Introduction

DeepSeek Harness (DSH) uses a plugin-based architecture and supports custom extensions. When building long-running agentic workflows, automatically preserving valuable content from conversations and avoiding context loss is a common challenge for developers. The dsh-motion-memory plugin provides a localized memory management solution through conversation tracking and periodic summarization.

What Is This

This is a memory management plugin designed for DeepSeek Harness (DSH). It is maintained by li3-feng2-jie2 and released under the MIT license. It automatically preserves content from conversations that is worth keeping into local memory documents, maintaining a long-term memory that “understands you better over time” through conversation tracking and periodic summarization. It supports local storage and local models end-to-end.

Core Features

The plugin mainly includes the following capabilities:

  • Four-layer memory structure: Divides memory into four layers—current/active, periodic, event, and raw conversation—with layer-by-layer traceability.
  • Conversation tracking: After each conversation turn, automatically pushes content into memory, including accumulated lessons learned.
  • Periodic summarization: Periodically consolidates and compresses active or important content to prevent unbounded memory growth.
  • Keyword deduplication and routing: Automatically checks for duplicates before writing, updates the same entity, and disambiguates different entities.
  • Citation jump navigation + read-only protection: Links in the content support traceable navigation; event and periodic memories are read-only by default, helping isolate contaminated content.
  • Memory panel enhancements: Supports filtering keywords by agent, categorizing by tags, and maintaining entries from the keyword library.
  • First-round overview injection: Injects items such as the user profile, work summary, and active keywords into the required memory area.
  • Two-stage updates: Downloaded updates are placed in a staging area only and are automatically activated after a restart, without interrupting operation.
  • Model-less fallback: Can be used without a configured model, with support for accumulating user message references.

Installation and Enabling

The plugin is hard-paired with DSH, so version compatibility must be confirmed before installation. The exact pairing for this version, v0.5.1, is DSH 0.2.0-rc.2.

Use the following command to install (it has not been published to npm, so npm package installation is currently unavailable):

dsh plugin --profile <profile> add github:li3-feng2-jie2/dsh-motion-memory

The installer automatically handles the cordis.patch.yml configuration, so manual profile editing is not required. DSH must be restarted after installation.

Typical Usage

Enabling the memory feature typically requires three steps:

  1. Configure a model: Go to the Settings page > “Memory Manager” and select a local model (a 7B–9B model loaded with LM Studio or Ollama is recommended).
  2. Enable tracking: Go to the Settings page > “Conversation Tracking”, enable the feature, and set the interval to 0 (triggers every turn).
  3. View memory: In the “Memory” panel on the conversation page, view turn summaries, keywords, and periodic summaries.

Use the following command on the conversation page to check the current status:

memory cmd=status

Use Cases and Notes

Use cases: Suitable for self-hosted AI workflow developers who need long-term memory and prioritize data privacy and controllability.

Notes:
* Hard version pairing: The plugin is hard-paired with DSH. Different versions may cause mount failures. Use dsh-motion-memory v0.5.1 strictly paired with DSH 0.2.0-rc.2.
* Model selection: A local model is recommended. Paid models can have high invocation costs due to low cache hit rates.
* Package manager limitations: Plugins installed via the package manager through “Add Plugin” are not supported for in-place activation. Use the DSH plugin page or the dsh plugin update command for upgrades.

Summary

dsh-motion-memory provides a structured local memory solution. Through its four-layer architecture and automated tracking and summarization, it helps DSH users build long-term memory. All data is stored locally, and when used with local models, it balances flexibility with cost control.

Project URL: https://github.com/li3-feng2-jie2/dsh-motion-memory