Preface

Agents often face the difficulty of “no memory” in each conversation and must understand the user context from scratch. Common memory solutions on the market either store raw conversations in a vector database, resulting in signal drowning, or allow the model to write arbitrary Key-Value entries, causing hallucination contamination. dsh-daoing-memory addresses these issues and provides a persistent, self-evolving, and auditable memory system for DeepSeek Harness (DSH) agents.

What It Is

This is a DSH plugin maintained by daoing. Its core value lies in establishing trust through an “earning” mechanism and converting raw conversations into reusable semantic memory. It separates semantic memory (user facts and concerns) from experiential memory (lifecycle knowledge), and records all changes through an append-only audit ledger to ensure the integrity and traceability of memory.

Core Features

This plugin provides full-chain tools from recording and extraction to recall and audit:

  • Notes and Extraction: Use memory_fact to append raw conversation notes; use memory_extract to distill notes into durable facts and concerns (such as decisions, commitments, questions, etc.).
  • Experience Lifecycle: Includes tools such as memory_ingest, memory_report, memory_revise, memory_refine, and memory_verify, which manage the trust-level promotion and correction process for experiences.
  • Relevance Recall: Through memory_recall, the agent can retrieve relevant content from the shared experience library based on keywords or context.
  • Audit and Cleanup: memory_ledger and memory_verify are used to query the append-only audit ledger; memory_consolidate is used for periodic compaction and storage cleaning.
  • Profile Injection: Injects a concise user snapshot (core facts and concerns) into the system prompt, enabling the agent to understand the user background without asking.
  • Workbench and Skill: Provides a browser sidebar (Workbench UI) for human review and editing; includes a built-in memory-extraction skill that guides the agent on when and how to extract high-quality memory.

Installation and Activation

After installing the plugin, restart the DSH process and manually place the skill file.

  1. Install the plugin:
    dsh plugin --profile web add dsh-daoing-memory
  1. Install the skill:
    node node_modules/dsh-daoing-memory/scripts/install-skill.mjs
  1. Restart the service:
    After installation, restart DSH to load the plugin and injection mechanisms. After restarting, a “Memory” button will appear at the bottom of the left sidebar for opening the Workbench.

Typical Usage

After the plugin is installed, the agent uses the memory_* tool series to collaborate with human users and build memory:

  1. Record raw notes: During a conversation, the agent uses memory_fact to append raw notes for the current session.
  2. Distill memory: At natural pause points, the agent runs memory_extract, combined with the memory-extraction skill, to extract facts and concerns from the notes.
  3. Retrieve and apply: In subsequent conversations, memory_recall automatically retrieves relevant experiences; meanwhile, the system prompt already includes a concise user Profile, eliminating the need for repeated questions.
  4. Feedback and refinement: Call memory_report when an experience is confirmed to be useful, or call memory_revise when an error is detected.

Use Cases and Notes

This plugin is suitable for DSH agent scenarios that require long-term user profile maintenance, complex task workflows, and a high level of data accuracy.

Notes:
* Memory must be “earned”: All experiences start as low-trust “candidates” and can only achieve high trust after being validated through real-world usage; they are not granted out of thin air.
* Human involvement: Although the plugin runs automatically, human users must remain engaged in reviewing, correcting, and deleting through the browser Workbench. Memory is a shared artifact, not a black box.
* Permissions and source code: The plugin runs with the permissions of the current DSH process. It is recommended to inspect the source code and license (MIT) before installation.

Summary

dsh-daoing-memory separates memory into semantic and experiential dimensions and, combined with an audit ledger and Workbench, solves the problems of agent memory contamination and context loss. Developers can refer to the GitHub repository for the complete architecture or visit the community directory for more details.