Introduction

In the DeepSeek Harness (DSH) plugin ecosystem, memory modules often require integration with third-party frameworks or rely on external APIs. memory-eternal solves this problem. It is an in-house plugin that does not depend on any existing memory framework and directly consolidates conversations into local Markdown.

What Is This

This is a DeepSeek Harness memory plugin authored by EternalNight996. Its core function is to automatically generate knowledge cards when a conversation ends and save them to a local Markdown Vault. It uses SQLite storage and in-house CJK retrieval, and provides a graphical interface and a knowledge graph.

Core Features

  • Automatic knowledge-card consolidation: Automatically generates knowledge cards to a local Markdown Vault when the conversation ends.
  • Retrieval and deduplication: In-house deduplication and CJK retrieval, with on-demand recall supported via the memory_recall tool.
  • Knowledge graph: Built-in knowledge graph, with a review center and a recycle center.
  • Data persistence: No third-party memory frameworks; uses SQLite for persistent storage.
  • Graphical interface: The “Memory” button at the bottom of the sidebar or the DSH settings page provides access.
  • Multiple knowledge libraries: Supports project-based memory isolation.
  • Ecosystem compatibility: Compatible with Claude Code / Codex / Cursor.

Installation and Enablement

  1. Open a terminal and run the installation command.
  2. Restart the DSH host (configuration changes take effect after a restart).
dsh plugin --profile web add memory-eternal

Typical Usage

  • Recall memory: dsh-memory recall "数据库选型"
  • Manual capture: dsh-memory capture "重要结论..."
  • Mine historical sessions: dsh-memory sweep ~/.claude/projects
  • Agent invocation: When an Agent needs history, it calls the memory_recall tool.

Use Cases and Notes

  • Suitable for developers who need to maintain a long-term project knowledge base.
  • Important limitation: It prioritizes, and only guarantees support for, the official DeepSeek Harness desktop edition (profile desktop), and is compatible with the web profile. The author’s in-house dsh-desktop is no longer maintained.
  • Resource consumption: Enabling “knowledge card distillation” or “semantic deduplication powered by AI” incurs AI invocation costs.
  • Protocol nature: MCP is a protocol, not a resident service.

Short Conclusion

For developers who need to build a long-lived memory library, memory-eternal provides an end-to-end in-house solution from storage to retrieval. View the directory page or GitHub for more details.