Introduction

When developing agents using DeepSeek Harness (hereinafter referred to as DSH), the philosophy that “everything is a plugin” implies numerous choices for memory capabilities. Hindsight, MemOS, and Graph Memory in the compatibility table fall into this category. The problem is that each memory plugin comes with its own storage and retrieval algorithms, and the methods to connect to them differ. To combine them in the same environment, you have to adapt interfaces one by one and maintain data separately. Changing one implementation requires touching the entire integration layer.

dsh-patchouli addresses exactly this issue: it separates “data access” from “algorithm implementation” and places a unified memory and knowledge service in between. The plugin is introduced below.

What is it?

The positioning of memorax-ai/dsh-patchouli is as a local memory and knowledge hub for DeepSeek Harness. The current version is 0.1.6 with an MIT license. The repository is hosted under the memorax-ai organization, and the author field in package.json is marked as CH4ACKO3.

It exposes only one unified update / retrieve / subscribe service to the outside. Internally, it works with several components: the Agent Loop connector provides configurable hooks and model tools; memory and knowledge implementations are pluggable, supporting both local and remote options; the underlying layer is a transactional Rust backend that supports SQLite and remote providers, responsible for persistence.

By the way, the name pays homage to “Patchouli Knowledge” from the Touhou Project series, as well as the widely known Patchouli mod for Minecraft.

Core Features

According to the README description, capabilities are focused on services, retrieval, access, and storage:

  • Common Memory Service: Supports provider filters, user routing strategies, and source tracing;
  • Consumer Aggregation: Keeps native plugin results unchanged, supports optional specialized model evidence selection;
  • Incremental Retrieval Stream: Includes backpressure control and final complete aggregation;
  • Agent Loop Connector: Provides configurable hooks and model tools;
  • Pluggable local or remote memory and knowledge implementations;
  • Image and workspace file ingestion, hosted as typed Artifacts;
  • Persistent subscriptions, and a transactional Rust backend supporting SQLite and remote providers.

Installation and Enablement

First, confirm the runtime requirements: Node.js ^22.19.0 || >=24, pnpm 11, and a DeepSeek Harness runtime compatible with 0.1.0-rc.6. Then follow these three steps:

  1. Execute the installation command:
dsh plugin --profile web add dsh-patchouli
  1. Use dump-config to check if the plugin is registered successfully:
dsh --profile web --dump-config

The output should list patchouli, patchouli-storage, and the connector plugins. Seeing these three categories indicates the installation is in place.

  1. Register at least one compatible memory or knowledge plugin to handle the routed update, retrieve, and subscribe calls. Without this layer, there is no handler for routed calls.

What Happens on First Run

The plugin comes with the patchouli-db dependency. On first use, it will download a verified daemon binary from the GitHub Release of the same version and initialize the default local database directory; the bundled DSH profile enables the storage client by default, connecting and starting the local daemon as needed.

For configuration and platform-specific details, see the Getting Started guide: https://memorax-ai.github.io/dsh-patchouli/getting-started

Compatible Memory and Knowledge Plugins

Compatibility falls into two categories: Official means the upstream plugin directly registers the patchouli service; Patch means it is adapted for a specific fixed precise version by GOOJFC via dsh-harmony. The dsh-harmony repository used for adaptation can be found at https://github.com/memorax-ai/dsh-harmony, and the plugin repository also marks it as Powered by dsh-harmony.

The compatibility table currently covers 11 plugins: MemoraX Code, OpenViking, Hindsight, MemOS, Mneme, Mnemon, Memory Gate, Lingshu Memory, Graph Memory, Engramory, and Memory Evolve. Each corresponds to a fixed test version, and the specific versions are subject to the compatibility table in the repository. For all plugins in the table, the Patchouli DB column is “No — plugin-managed,” meaning these plugins’ databases are not taken over by Patchouli.

A notable detail: The compatibility patch for Engramory preserves its index guard and sets the registerSkill: false option as per its documentation; during the adapter’s active period, Patchouli is the only automatic recall path.

Suitable Scenarios and Precautions

Suitable scenarios:

  1. Environments that use multiple memory/knowledge plugins simultaneously and wish for unified access and routing;
  2. Deployments that require local persistence (SQLite) or connection to remote providers;
  3. Developers who want to integrate their own memory implementations, simply by interfacing with the update / retrieve / subscribe set of calls.

Precautions before use:

  1. The plugin runs with the permissions of the current DSH process; check the source code and license before installing (this project is MIT);
  2. Patch-type compatibility bindings fix the plugin version; verify the repository’s compatibility table before upgrading the upstream plugin;
  3. The first run downloads the daemon binary from a GitHub Release (with checksum verification); offline environments need to be aware of this step.

Conclusion

After the steps above, Patchouli’s role in DSH is clear: the connector receives data, memory and knowledge plugins perform the algorithms, and the Rust backend manages storage. This is its most direct value for environments that want to combine multiple memory plugins or switch implementations without touching the integration layer.

  • GitHub Repository: https://github.com/memorax-ai/dsh-patchouli
  • Community Directory Page: https://www.skillhub.cn/plugins/memorax-ai/dsh-patchouli

Additional note: The community directory is an independent site and has no official affiliation with DeepSeek or Fantom (Huanfang).