Introduction

In DeepSeek Harness (DSH), allowing the model to directly handle long-term memory, logging, or database permissions can introduce complexity. dsh-hypatia solves this by integrating the Hypatia knowledge graph into the DSH plugin system. It runs the Hypatia CLI as a host process, letting the model focus on proposing memories, while the plugin itself handles database orchestration, permissions, and retries.

Plugin Positioning

dsh-hypatia is a memory plugin. It is maintained by user tkliuxing. It provides long-term memory for DSH, backed by the Hypatia knowledge graph. The plugin runs Hypatia in host-side code; the model is only responsible for proposing what is worth remembering.

Core Features

The plugin provides the following capabilities:

  1. Recall within the same request: Retrieves and appends relevant project memories within fixed time and size budgets.
  2. Precise project scope: Memories belong to a single project, derived from the standard workspace path; cross-project leakage is prevented by the host-side ledger, rather than relying on content tags.
  3. Verified writes: Every write is read back and compared, and is marked as stored only after success is confirmed.
  4. Two-stage forgetting: The exact entry is shown before deletion, and cleanup status is reported honestly.
  5. No Bash required: Memory works in read-only and write sessions because the plugin never asks the model to execute shell commands.
  6. Knowledge graph query skills: Provides the ability to query the knowledge graph.
  7. Automatic conversation memory bridging: Automatically handles bridging between conversation and memory.

Installation and Dependencies

Before installation, ensure the system meets the following conditions:

  • Hypatia CLI: Must be available in PATH. The plugin depends on the hypatia 0.1.4 CLI contract.
  • Node.js: Version 22.5 or later (the control ledger uses node:sqlite).

The installation command is as follows:

dsh plugin --profile web add @tkliuxing/dsh-hypatia

Configuration

After installation, configuration is usually done in the cordis line. The key configuration items are:

  • memory.preset: Memory preset options, including disabled, read-only-recall, standard (default), and full.
  • memory.authorization: Memory authorization is independent from the DSH file sandbox.
  • recall: Controls recall behavior, including the enable switch, deadline, maximum number of results, and so on.
  • ingest: Controls summary ingestion.
  • reconcile: Controls reconciliation operations.

Usage Examples

The plugin provides the following tools for agents:

  • Memory storage:
    • “remember: this project forbids eval”
  • Memory retrieval:
    • “what do we know about the retry policy?”
  • Memory forgetting:
    • “forget what you know about the old API” (preview first, then confirm deletion)
  • Status check:
    • “did that actually save?”
    • “settle whatever is still unverified”
  • Knowledge graph management: Managed by driving the CLI directly through the hypatia skill; requires danger-full-access permission.

Limitations and Notes

  • Single-process limitation: Only one process can run at a time. Concurrent calls will fail due to file locking.
  • Version requirement: A Hypatia CLI compatible with the 0.1.4 contract is required.
  • License: MIT License.

Ecosystem and Resources

The DeepSeek Harness plugin ecosystem emphasizes “everything is a plugin.” This plugin follows community directory specifications and is not affiliated with DeepSeek official.

  • GitHub repository: https://github.com/tkliuxing/dsh-hypatia
  • Community directory: https://www.skillhub.cn/plugins/tkliuxing/dsh-hypatia