Introduction

In the daily use of DeepSeek Harness (DSH), agents often fail to remember project conventions or personal preferences from previous conversations, causing repeated confirmations. The dsh-behuman plugin solves this problem through long-term memory and skill-evolution mechanisms.

Plugin Overview

dsh-behuman is a plugin for DeepSeek Harness that provides long-term memory and self-evolving skills, maintained by goodddGrades. It enables agents to remember user preferences, corrections, and project state across sessions, and automatically turns successful methods into reusable skills.

Core Features

  1. Cross-Session Memory Management: The plugin remembers user preferences, project conventions, and corrections, and recalls this information in new sessions through context, without manual management.
  2. Automatic Skill Evolution: When a task involves more than 5 tool calls and is successfully completed, the plugin automatically writes the method to the skill library for direct reuse in similar future tasks.
  3. Zero-Configuration Operation: After installation, no manual memory-management configuration is required; it runs automatically in the background.
  4. Pure File Storage: Memory and skills are stored as plain Markdown files on disk, without using a vector database.
  5. Automatic Review Mechanism: After 10 tool calls without any written content, the plugin automatically runs a review to fill in missing information. This is the only step that consumes extra model calls.
  6. No Retrieval Design: The directory exists only as context; no retrieval or search is performed, and information is passed directly through context.

Installation and Enablement

According to DSH’s profile mechanism, it must be installed into the currently running profile. If running dsh web, execute the following command:

dsh plugin --profile web add @goodddgrades/dsh-behuman

If running dsh tui, execute:

dsh plugin --profile tui add @goodddgrades/dsh-behuman

No additional enablement is required after installation; use it directly.

Typical Usage

  • Writing Memory: When you mention preferences for answer style, project directory structure, or environment configuration, or when you correct an agent error and the correction succeeds, the plugin automatically writes the relevant facts to the .dsh/memory directory.
  • Writing Skills: When the agent completes a task that includes more than 5 tool calls and the process is validated as successful, it encapsulates the workflow into a skill file.
  • Facts vs. Instructions: The plugin prioritizes recording facts (such as “the user prefers short answers”) rather than instructions (such as “always answer briefly”), to avoid instructions overriding the user’s true intent in later sessions.

Configuration and Storage

The plugin supports configuration, with default paths as follows:

  • Memory directory: dir (default .dsh/memory)
  • Skills directory: skillsDir (default .agents/skills)
  • Review interval: nudgeInterval (default 10, meaning review is triggered after 10 tool calls without writing)
  • Review backend: reviewBackend (default spawn)

Memory is stored as Markdown files, and skills are stored in SKILL.md format. Handwritten skills will not be overwritten. The directory grows linearly; after several hundred lines, efficiency may decrease, at which point consider using a searchable index instead of a database.

Use Cases and Cautions

  • Use cases: Suitable for DSH projects requiring long-term maintenance, multi-person collaboration, or high context coherence.
  • Note: The review step incurs additional model-call costs.

Summary

dsh-behuman gives DSH agents long-term memory and reduces repeated confirmations. For more details, see:
* Plugin directory: https://www.skillhub.cn/plugins/goodddGrades/dsh-behuman
* GitHub: https://github.com/goodddGrades/dsh-behuman