dsh-plugin-memory

Every session in DeepSeek Harness starts from a blank slate. The project context you explained last week, your working preferences, the task you were halfway through—all gone the moment you open a new window. dsh-plugin-memory solves this by installing a five-layer persistent memory inside the harness: a user profile, a per-project semantic index with topic files, and append-only daily logs, all stored locally on disk with no external service required. The agent stops being a one-shot dialogue machine and becomes a working partner that actually remembers you.

Two pipelines drive the core capabilities. On the read side, every step selects relevant topic files based on the current query—semantic ranking when an LLM is configured, keyword scoring as a fallback—and injects them into context, with automatic deduplication within the same session and a hard-clamped cold-start index to keep tokens cheap. On the write side, once a session goes idle the plugin debounces a 60-second timer, scans the most recent 40 events, and asks the LLM to distill new facts into topic files and index lines. Existing memories are never overwritten; if the model is unavailable the step degrades silently. The agent also has six memory_* tools for reading, writing, searching, and forgetting memories, plus merging and rotating the user profile across four fixed sections.

Built for developers who work long-term in Harness across multiple projects and want the agent to accumulate context instead of restarting from zero every time. Basic injection and keyword matching work without any LLM route; plug in a model for the full experience, including auto-extraction and semantic ranking.