dsh-ov-memory

OpenViking is a solid persistent-memory backend, but on its own it never speaks up during a conversation. dsh-ov-memory is the missing bridge: before each turn, it retrieves the most relevant fragments from OpenViking and injects them into the context window, with server-side deduplication and token-budget control. After the exchange, session messages are mirrored into the dsh- conversation stream in real time so the backend can asynchronously distill them into long-term memory. The whole pipeline runs with zero extra npm dependencies, reusing the peer packages already installed by DeepSeek Harness.

Two resilience paths keep the integration honest under real-world conditions. When OpenViking is unreachable, every write is persisted to a local outbox and replays idempotently on the next session start, with content-hashed dedup, a retry ceiling, and TTL-based cleanup. The /memlearn slash command lets you crystallize a single insight into long-term memory on the spot: secrets are redacted first, the text is semantically matched against existing entries and merged into the closest one, and if the similarity score falls below the threshold the plugin honestly refuses rather than force a poor fit. Under the hood, a purpose-built minimal MCP implementation (stdio server plus streamable-HTTP upstream, no third-party MCP SDK) exposes every OpenViking tool as mcp__openviking__* names, auto-syncing when the server tool table changes, while a viking:// URI guard stops local fs and shell tools from treating virtual paths as real files.

This plugin is for developers who are already running DeepSeek Harness alongside OpenViking and want to fold persistent recall and session capture into the conversation loop without stitching together their own glue code. Requirements are modest: an accessible OpenViking server (local or remote), Node.js ^22.19.0 or >=24, and a DSH release on the 0.1.x line.