AI Agent Hub
Back to plugins
🧠

dsh-plugin-om

Memory Updated 2026.09.05

Run the following command in DeepSeek Harness:

dsh plugin install FanetheDivine/dsh-plugin-om

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install FanetheDivine/dsh-plugin-om in your DeepSeek Harness terminal to install this plugin; the source repository is https://github.com/FanetheDivine/dsh-plugin-om .

About this plugin

Long-running DSH sessions fill the context window quickly, and naive truncation throws away information you still need. dsh-plugin-om adopts an Observational Memory strategy: before each agent step it automatically compresses earlier messages into summary blocks, and once those blocks exceed a second threshold they are merged into a single tighter digest, reducing token pressure while preserving semantic content.

Beyond auto-compression, the plugin ships two recall tools: recall retrieves original messages by index range including image attachments, and recall-semantic runs a local embedding model to perform semantic search across all full messages so content buried inside summaries can still be surfaced. Abstract parsing uses a tag-anchored extraction pipeline with a fuzzy-rebuild fallback, and when core services misbehave the compressor degrades gracefully, reporting a single warning event rendered as an expandable row in the browser client without blocking the session.

If you rely on DSH for multi-turn agent workflows or long-chain reasoning and want tighter context management than the built-in compaction-basic while keeping the ability to trace back to original messages at any time, this plugin is a strong fit. Even with auto-compression disabled, the recall and recall-semantic tools remain fully functional, making it suitable for everything from solo long sessions to multi-agent coordination pipelines.

Use Cases

  • Automatically compress long session history to control token usage
  • Inspect original messages by index even after compression
  • Search all full messages semantically using a local embedding model

Best For

  • DSH users who need fine-grained context management in multi-turn agents
  • Developers handling long-chain reasoning while retaining recall of original messages
  • Teams that want more control over summarization pacing than default compaction