DeepSeek Harness Plugin: Domain-Specific Long-Term Experiential Memory

The DeepSeek Harness (DSH) ecosystem is based on a plugin architecture, and agents often find it difficult to reuse experience across sessions during multi-turn interactions. marquez807/dsh-experience-memory is designed to solve this problem by providing DSH with domain-specific long-term experience memory.

Core value: Distinguish what matters, accumulate experience, discard expired items, correct errors, and automatically recall relevant experience when performing similar work again.

Installation and Activation¶

The plugin is installed via a tarball and supports running with zero configuration.

dsh plugin --profile <name> add /path/to/dsh-experience-memory.tgz

No additional configuration is required after installation. The default database is located at $DSH_HOME/experience-memory/memory.db.

Core Features¶

The plugin has no third-party dependencies and uses only Node.js built-in capabilities and SQLite.

• Low-overhead memory: Each conversation turn consumes a fixed 204 bytes to indicate that memory is available.
• Tools and commands:
  • Provides 5 model tools: memory_recall, memory_remember, memory_feedback, memory_forget, and memory_stats.
  • Provides 7 slash commands, such as /memory-status and /memory-preview.
• Evidence grading mechanism: The plugin grades memories based on the evidence source, including verified-tool (tool calls), verified-user (user’s original words), verified-file (file contents), and inferred (the default grade for unverified content). Only records with a high grade and matching relevance are injected.
• Automatic injection and maintenance: Automatically maintains the memory store and injects relevant experience according to relevance.

Usage Examples¶

Check the current memory store status:

/memory-status

Preview which memories will be injected in the current session:

/memory-preview 部署

Run the self-check tool to verify whether referenced files are missing:

node tools/provenance-audit.mjs

Notes¶

• Log confirmation: After restarting, check the startup logs to confirm that store <path> is the expected store path and record count.
• Data safety: Retirement is reversible; only executing purge truly deletes data.
• Decorator limitations: Markdown decorations are used only for diagnostics and do not affect the grading logic.
• File ownership: A record’s file ownership is resolved by the current working directory of the session at write time.

This plugin is suitable for DSH projects that need to accumulate domain knowledge during long-running operation. For more details, see: GitHub.