Foreword

DeepSeek Harness (DSH) is a developer preview tool that provides flexibility through a plugin-based architecture. In agent development, it is often difficult to distinguish between “whether the memory mechanism is truly effective” and “improvements in the model’s generalization ability.” dsh-memoryos aims to solve this problem by providing long-term project memory through an evidence-first strategy, along with strict baseline comparison capabilities.

What Is This

This is a plugin that provides evidence-first long-term project memory for DeepSeek Harness (DSH), powered by MemoryOS at the underlying level. It is a pure JavaScript Bundle maintained by tianhao8687, designed to add scoped project memory tools and precise Provider usage records to the DSH process, without altering MemoryOS’s persistence, retrieval, ground-truth resolution, or context compilation logic.

Core Features

  • Evidence-first long-term project memory: Focuses on controllable, verifiable agent evidence around memory claims.
  • Scoped project memory tools: Provide memory access targeted at a specific project.
  • Precise Provider usage collection: Accurately records input/output/cache usage, separated from estimated MemoryOS component attribution.
  • Strict no_memory baseline: Under the no_memory condition, the entire component—including control tools—is completely absent, retaining only the usage collector.
  • Control tools: Allow enabling or disabling MemoryOS in chat, with state persisted.
  • Context modes: Support compact, progressive, explanatory, and incremental context modes.
  • Explicit write configuration: Supports policies such as supersede, keep_both, and reject.
  • Read-only by default: Read-only by default; writes and controlled context eviction require explicit opt-in.

Installation and Activation

This Bundle is a DSH adapter, not a database. The local MemoryOS service must be started first.

  1. Start the MemoryOS service
    Run the following in the MemoryOS source code directory:
    python -m memoryos --data-dir ./data serve --port 8000 --no-open
  1. Install the DSH Bundle
    Use a release tag for reproducibility:
    dsh plugin --profile memoryos add github:tianhao8687/dsh-memoryos#v0.2.0
    dsh --profile memoryos --dump-config
  1. Configure environment variables
    Set the relevant variables when starting DSH (PowerShell example):
    $env:MEMORYOS_BASE_URL = 'http://127.0.0.1:8000'
    $env:MEMORYOS_AUTH_TOKEN = '<local-memoryos-token>'
    $env:MEMORYOS_CONDITION = 'msc_context_only'
    $env:MEMORYOS_BUDGET_TOKENS = '512'
    $env:MEMORYOS_MAX_CONTEXT_CALLS = '1'
    $env:MEMORYOS_RESPONSE_FORMAT = 'deepseek-compact'
    dsh --profile memoryos

Typical Usage

The plugin provides the ability to control MemoryOS through chat, without requiring desktop shortcuts or shell commands:

  • Enable/disable: Enter turn off OS or turn on OS. The model must call the memoryos_control tool.
  • Query status: Enter Is OS on now?
  • Install command:
    dsh plugin --profile memoryos add github:tianhao8687/dsh-memoryos#v0.2.0
  • Configure environment variables:
    $env:MEMORYOS_BASE_URL = 'http://127.0.0.1:8000'

Applicable Scenarios and Notes

  • Applicable scenarios: Developers who require strict evidence, baseline comparison, and fine-grained control over the scope of memory context.
  • Compatibility: Pinned to DeepSeek Harness 0.1.0-rc.5 (commit 47f943859bef60e4160492346772ded9b24f765a).
  • Architecture note: The Bundle is a DSH adapter, not a database. SQLite, migrations, retrieval, current ground truth, conflict relationships, and context compilation logic remain in MemoryOS.
  • Read-only by default: Read-only by default. Write evaluation and controlled context eviction require explicit opt-in.
  • Strict baseline: When MEMORYOS_CONDITION=no_memory, the entire component—including control tools—is completely absent, retaining only the usage collector.
  • Security note: The plugin runs with the current DSH process privileges. Please review the source code and license before installation.

Short Conclusion

dsh-memoryos helps developers validate the effectiveness of memory mechanisms by providing a strict no_memory baseline and precise tracing. For more details, refer to the official directory: https://www.skillhub.cn/plugins/tianhao8687/dsh-memoryos or the GitHub repository: https://github.com/tianhao8687/dsh-memoryos.