Introduction

The philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” An agent typically forgets everything between sessions, and memory tends to accumulate redundancy and noise during repeated use. The dsh-evolve plugin provides persistent memory and skill lifecycle management for DeepSeek Harness, aiming to solve the forgetting problem and prevent knowledge base bloat.

Plugin Positioning

dsh-evolve is a Self-evolving memory and skill lifecycle plugin developed by maintainer chenzheshushi-commits. It does not preset any viewpoints; it only provides mechanisms and rules. All learned content is stored locally and does not leave the environment.

Core Capabilities

The plugin includes the following core features:

  1. Cross-session memory and deterministic recall
    Provides structured records (fact / preference / decision, etc.) stored as JSON and Markdown mirrors. Uses a zero-token deterministic recall algorithm (combining bigram-Jaccard similarity and SQLite FTS5 BM25) to automatically inject relevant memory based on the current message without calling an embedding API.

  2. Tiered approval mechanism
    Memory generated by the model passes through a deterministic gateway that uses factors such as reversibility, conflicts, and overlap to decide whether to auto-confirm it or route it to a pending-review queue. This mechanism ignores the model-self-reported kind field to ensure the reliability of writes.

  3. Reinforcement learning and memory consolidation
    Repeated observations reinforce understanding rather than merely repeating content. By increasing observation counts and importance thresholds, combined with configuration parameters, high-frequency knowledge points are given higher weight.

  4. Skill lifecycle and in-place refinement
    High-value instructional content can be crystallized into SKILL.md files. New evidence is used to refine existing skills rather than creating duplicate files. Skills have three states: active / stale / archived, supporting in-place optimization and manual archiving.

  5. Bloat prevention and convergence
    Performs deduplication and cleanup for skills and memory. On the skill side, it detects similar content and merges it; on the memory side, it sets character budgets to prevent unbounded data growth.

  6. Background per-turn review
    After each conversation turn ends, the background process replays the current conversation snapshot through an independent LLM call to determine which content is worth remembering. This process is routed to a separate reviewModel and does not affect the main conversation flow.

  7. Web settings page
    Provides a visual configuration interface.

Installation and Activation

Installing this plugin requires the following prerequisites: Node.js >= 22.5.0 and DeepSeek Harness version >= 0.1.0-rc.7.

Add the plugin using the official installation command:

dsh plugin --profile web add github:chenzheshushi-commits/dsh-evolve

After installation, restart Harness. The tools will be detected at startup.

Typical Usage

Use the following commands:

  • Memory management

    • memory_remember: Store facts or preferences.
    • memory_recall: Recall relevant memories.
    • memory_confirm: Confirm a memory that requires review.
    • memory_confirm_batch: Confirm memories in batch.

  • Skill management

    • crystallize_skill: Create a SKILL.md file from marked lessons.
    • evolve_maintain: Perform maintenance cleanup of the knowledge base.

  • Configure the review model
    Set reviewModel in the configuration to route background reviews to a different model (for example, a cheaper or more capable model).

Technical Requirements and Notes

  • Environment dependencies

    • Requires Node.js >= 22.5.0 to use the built-in node:sqlite module for full-text search.
    • Requires DeepSeek Harness 0.1.0-rc.7 or higher.

  • Optional dependencies

    • git: Enables automatic checkpoint and rollback features.
    • tar: Enables pre-operation backups and the skill_rollback functionality.

  • Operating system
    The plugin is developed and tested on Linux and macOS; Windows environments have not been tested.

  • Data safety
    The plugin never deletes data. All change operations are reversible and can be restored via rollback or backup.

  • Plugin directory: https://www.skillhub.cn/plugins/chenzheshushi-commits/dsh-evolve
  • GitHub repository: https://github.com/chenzheshushi-commits/dsh-evolve