Introduction

When using DeepSeek Harness, models can easily lose context or repeat the same mistakes. Manually organizing experiences is time-consuming and prone to omissions. The dsh-dream plugin solves the problem of long-term memory and experience reuse for agents through automated extraction and structured storage.

What Is This

dsh-dream is a DeepSeek Harness plugin maintained by STARDUSTLC666. It is categorized under “Memory,” and its core function is to automatically organize candidate experiences from tasks and reuse them in related tasks after review. It also supports manual retrospectives and saving dream diaries.

Core Features

This plugin provides a complete workflow from extraction to reuse:
1. Browsing and Retrieval: Browse dream diaries, retrieve experiences, and view associated sources.
2. Automatic Extraction and Reuse: After a task is completed, automatically extract candidates, and automatically retrieve adopted experiences in relevant new tasks. Two independent toggles control these behaviors.
3. Consolidation Strategy: Consolidate entries after quiet periods, and preserve pending rounds when cooldown limits or quotas are insufficient. Usage and contributions can be controlled separately for each chat.
4. Review and Feedback: Supports direct review, conflict handling, and recording independent usage feedback.
5. Batch Management: Review in batches by project and topic. Pause reuse for expired facts, and restore them after re-verification.
6. Rule Writing: Preview rule differences, write them to AGENTS.md, and support rollback.

Installation and Enabling

In the desktop version of DSH, use the following command to install the plugin:

dsh plugin --profile desktop add @stardustlc/dsh-dream

After installation, restart DSH. This plugin is licensed under the MIT license.

Typical Usage

Normal use of DSH can trigger the automatic consolidation workflow. When a completed task round shows a clear preference, a correction, or at least two non-Dream tool calls, candidate experiences will be organized within the budget. Check the pending review prompt under Settings → Dream Diaries, verify the sources, and adopt them. Subsequent related tasks will reuse the adopted experiences automatically.

You can also generate a diary manually, for example by saying: “Review the recent tasks and record the experiences worth retaining.”

Applicable Scenarios and Notes

  • Review Mechanism: New experiences are always pending review first. Expand them to verify the original wording, conditions, and exceptions before adoption. Facts are marked for re-verification by default after 30 days, while stable preferences do not have a default expiration time.
  • Retrieval Behavior: Retrieval only indicates reference retrieval, not re-verification. It does not automatically adopt experiences or write project rules. Rule writing still requires previewing and applying changes.
  • Language and Privacy: The interface follows DSH’s language in both Chinese and English, while diary and experience content remains in its original language. By default, it uses local DSH sessions and diary directories, and supports privacy masking.
  • Interface Compatibility: Hosts without an automatic interface retain tool and skill entry points.

Conclusion

By turning experience extraction, review, and reuse into a structured workflow, dsh-dream helps DSH users build long-term memory. To learn more details or submit feedback, visit the GitHub repository or check the plugin directory.