DeepSeek Harness / OpenViking
DeepSeek Harness PluginOpenViking
dsh-memory-plugin
MCP bridge
Developer preview
Keep working preferences available between tasks
Keep a repository’s working preferences available when your AI coding agent returns to the next task. The OpenViking DeepSeek Harness plugin connects DSH to a memory server, recalls relevant context before model steps, and captures conversation events for later reuse.
Start with the OpenViking setup guide, or read how the DSH plugin architecture handles storage to understand the extension points.
The Memory Server Behind dsh-memory-plugin
Connect the same DSH profile to a reachable OpenViking server before asking it to remember project conventions. The installed package, @openviking/dsh-memory-plugin, bridges the agent’s lifecycle and tools to that server.
The integration supports three workflows:
| Your work | Plugin action | Result available to DSH |
|---|---|---|
| Start a session or model step | Load profile context and retrieve relevant memories | Additional context for the current task |
| Continue the conversation | Capture eligible session events and commit accumulated content | Material for later memory retrieval |
| Request a memory operation | Route through the OpenViking MCP bridge | Server-backed reads, searches, or writes |
A server supporting the viking://~ home alias is required. Keep this memory workflow separate from finding evidence in an existing document collection; WeKnora knowledge-base retrieval covers that neighboring task.
Auto-Recall and Profile Injection Before Each Step
Give the agent a concrete task that can benefit from a previously recorded convention. At the pre-step hook, the plugin implementation requests profile context and recall concurrently, then appends available context as durable plugin messages.
Try these suggested inputs after setup:
Suggested prompts
- “Use this repository’s remembered test conventions when planning the parser change.”
- “Find the previous decision about error-message formatting and explain how it applies here.”
- “Summarize the project preferences available for this task in a short checklist.”
The last prompt requests a checklist you can inspect before editing code. Relevant stored material determines what recall returns; the prompt itself does not create evidence that a convention exists.
For a focused repository workflow, choose OPENVIKING_RECALL_PEER_SCOPE=actor. The default all scope also searches the user’s other peers. Repository identity normally follows the normalized Git origin, while OPENVIKING_PEER_ID lets you supply an explicit identity.
Session Capture and viking:// URIs, Explained
Continue working after recording a useful preference, then revisit it in a later task. User and assistant messages enter the capture stream; tool-result capture is optional. Failed writes enter a pending queue for replay at a later session start.
Think of the capture sequence as:
conversation event -> capture -> commit -> later recall
These stages are distinct. A preference mentioned in conversation must reach the server’s committed memory before retrieval can use it.
Treat a returned viking:// address as a server resource. The URI guard and capture controls redirect filesystem or shell attempts toward the bridged OpenViking tools. A Viking address is therefore used through the memory interface, rather than opened as a local project file.
Setting Up Your Own OpenViking Server
Prepare Node.js ^22.19.0 or >=24, DSH 0.1.0-rc.6 or newer within 0.1.x, and an installed OpenViking server.
# 1. Initialize the server; the wizard prepares its configuration openviking-server init # 2. Check configuration and model access, then start the server openviking-server doctor openviking-server # 3. Configure the plugin's endpoint and credentials via OPENVIKING_* vars # or edit ~/.openviking/ovcli.conf # 4. Add the package to your DSH profile dsh plugin --profile web add @openviking/dsh-memory-plugin # 5. Inspect the dumped config for the openviking-memory group dsh --profile web --dump-config
Open that profile and try a recall prompt once the openviking-memory group appears in the dump. The server deployment guide distinguishes liveness at /health from readiness at /ready. Use readiness when a running process still cannot serve the expected memory workflow.
For the surrounding configuration model, continue with EvoMap’s DSH plugin architecture guide.
Compatibility
DeepSeek Harness is currently in developer preview.
| OpenViking bundle | 0.3.1 |
|---|---|
| Source check date | 2026-09-08 |
| DSH peer range | >=0.1.0-rc.6 <0.2.0 |
| Minimum-contract fixtures | Pinned to rc.6 |
These declarations were inspected from the OpenViking bundle manifest, with no local integration run performed.
OpenViking: frequently asked questions
Can I recall memory without saving this conversation?
Set syncTurns: false in the plugin configuration. Profile injection and recall continue, while captured turns, commits, and queued-write replay stop; pending writes remain queued for a session that permits writing.
How do I keep another repository’s memories out?
Set OPENVIKING_RECALL_PEER_SCOPE=actor to restrict retrieval to the actor peer. Use an explicit OPENVIKING_PEER_ID when your workspace needs a deliberately chosen identity.
Does installing the plugin also provide the memory server?
A reachable OpenViking server is a separate prerequisite. The package adds the DSH integration; point it at the server you configured before checking recall.
Are tool results captured automatically?
Tool-result capture is optional and controlled by the integration configuration. If your priority is preserving a detailed investigation trail, review the separate EverOS trace-capture workflow and its data boundaries.
Other DeepSeek Harness memory & data plugins
MemOS
dsh-memory-plugin alternative
Compare MemOS’s memory integration against OpenViking’s approach to recall and profile injection.
EverOS
setup & permissions
See how EverOS handles trace capture and permissions for detailed investigation trails.
WeKnora
document tools
Use WeKnora for knowledge-base retrieval, a neighboring but distinct task from memory recall.
