Introduction¶

When developing Agents with DeepSeek Harness (DSH), the conversational context typically vanishes when the window closes. In the next new session, the model often doesn’t remember projects discussed last week, coding preferences, or decisions made, forcing developers to repeatedly explain the background.

The DSH ecosystem’s philosophy is “everything is a plugin” – capabilities are extended through plugins, and community directories are maintained by independent sites like SkillHub, with no official affiliation to DeepSeek or High-Flyer. Among the “memory” plugins, dsh-mneme focuses on cross-session persistence and background processing: it writes key information from conversations to local storage, automatically recalls relevant content in new sessions, and performs deduplication, merging, and archiving during idle times.

Below, we introduce this plugin’s positioning, capabilities, and installation method. The factual basis is the GitHub repository and the npm package @modusensus/dsh-mneme (current version 0.7.1, MIT license).

What It Is¶

dsh-mneme is maintained by modusensus and is a “memory” plugin for DSH. The name comes from the Greek memory goddess Mnemosyne (Μνήμη); the README refers to the background consolidation mechanism as autoDream, corresponding to memory maintenance during the “sleep” phase.

In one sentence: it provides persistent cross-session memory for Agents – automatically recording key information during conversations and injecting relevant memories into new sessions; data is stored locally in Markdown format, processed locally by default, and does not rely on remote services.

Compared to the approach of “just stuffing the context window in a single conversation,” it stores memory in structured storage (entities, attributes, relationships, and timelines), and combines semantic retrieval and graph recall to gradually refine the memory base with use, rather than simply piling up text.

Core Features¶

The following capabilities are derived from the README and package.json descriptions, summarized along the usage flow.

Cross-Session Memory and Recall¶

• During conversations, the AI automatically records key information; when starting a new conversation, it injects memories relevant to the current topic.
• Starting from v0.5.0, it supports recall fusion: a combination of BM25, graph, and “hot memory” retrieval (README roadmap).
• v0.7.0 introduces a self-evolution mechanism: heat power-law decay, dual protection for sleep, and entity heat projection – keeping frequently used memories active while letting less common content fade over time.

Automatic Organization (Sleep Mode)¶

• v0.4.0 introduced Sleep Mode: performs four-stage deep maintenance (deduplication, merging, archiving, etc.) when the system is idle.
• The README states the memory base “becomes more refined with use,” corresponding to the autoDream self-consolidation description in package.json.

Structured Entities (Optional)¶

• Starting from v0.3.0, it has “memory genes”: three tables for entities, attributes, and relations, plus a timeline.
• The configuration option entityExtractionEnabled defaults to false; after changing it to true in the DSH settings panel → Memory Base Settings, it allows the AI to automatically extract structured information. Detailed design can be found in dsh-mneme/docs/ENTITIES.md in the repository.

Deleting a Conversation ≠ Deleting Memory¶

• v0.6.0 supports session lifecycle: sessionLifecycleEnabled defaults to false; after changing it to true, deleting the chat window will not clear already archived memory (behavior is configurable).

Local and Privacy¶

• Memory is stored locally as Markdown files, human-readable and manually editable.
• The README promises: no data is uploaded to servers; zero network dependencies by default, no API Key; no telemetry or remote logs.
• For fully offline semantic capabilities, change embedProvider from the default openai to local; local model deployment instructions are in dsh-mneme/docs/LOCAL_MODEL.md.

Engineering Status¶

• Requires Node 24+ (depends on node:sqlite).
• CI indicates 757 tests pass; GitHub shows approximately 43 stars and 3 forks (as of the information capture time).

Installation and Activation¶

The plugin targets the DSH web profile. After installation, it registers the bundle layer and requires restarting the DSH process.

dsh plugin --profile web add @modusensus/dsh-mneme
dsh web

It is ready to use immediately after installation. Optional configurations can be modified in the DSH settings panel → Memory Base Settings; complete instructions are in the repository’s dsh-mneme/docs/CONFIG.md.

Security Note: The plugin runs with the current DSH process permissions and can read/write the local memory directory. Before installation, it is recommended to browse the source repository and the MIT license to confirm it meets your environment and compliance requirements.

Typical Use Cases¶

1. Install the Plugin and Start the Web Interface¶

After installation using the commands from the previous section, run dsh web and adjust the memory base options in the settings as needed.

2. Long-Term Project Collaboration¶

Discuss project requirements on Monday and close the window; when starting a new conversation on Wednesday, the Agent can continue the context based on stored memories (README examples: blog refactoring, technology selection with Astro, etc.), without needing to repeat everything from scratch.

3. Set Coding Preferences Once¶

Explain coding habits or team standards in a conversation; after the plugin archives them, subsequent sessions will automatically incorporate them, reducing the need to repeat explanations each time.

4. Retrieval After Organizing Materials¶

After organizing a large amount of material through conversations and archiving it; even after closing the window, it can be retrieved through semantic retrieval and graph recall, rather than being destroyed with the session.

5. Local Development and Verification (Maintainer Workflow)¶

If you need to modify the plugin or run tests, the README provides the following commands:

cd dsh-mneme
npm install
npm test          # 757 tests
npm run stress    # Three-axis stress test
npm run sync      # src → lib synchronization

Applicable Scenarios and Notes¶

Who It’s For

• Developers who frequently use the DSH Web interface for project development and material organization, wanting the Agent to remember preferences and decisions.
• Users who value data staying local, editable Markdown, and optional full offline capability.
• Scenarios requiring automatic memory organization with use, rather than manually maintaining a notes base.

Usage Notes

• The default embedProvider is openai; in strictly offline environments, be sure to change it to local and deploy the local embedding model according to LOCAL_MODEL.md.
• sessionLifecycleEnabled is disabled by default; if you want to retain memory after deleting a session, you need to manually enable it.
• The roadmap indicates that v0.8.0 (graph enhancement: interest drift, cross-workspace sharing) is still in planning, with specific capabilities subject to official release.
• Community directory page: SkillHub - modusensus/dsh-mneme; refer to it alongside the GitHub README.

Conclusion¶

dsh-mneme advances the DSH Agent from “amnesia within a single conversation” to a memory layer that is “cross-session, organizable, and locally controllable”: install the plugin, adjust three configurations as needed, and you can accumulate and recall project context locally. Further details are available in the GitHub repository and the SkillHub directory page.