dsh-layered-memory

Every time you open DeepSeek Harness, the agent starts from scratch - your writing style, business constraints, and unfinished context from the previous session are all gone. dsh-layered-memory distills conversation content into three tiers of long-term memory: a transient layer for current session highlights, an episodic layer storing one summary per session, and a semantic layer accumulating your habits, style, business needs, and long-term preferences. Install once, and subsequent sessions pick up where you left off automatically.

The core design centers on automation. The plugin captures user messages in real time via an agent/pre-step hook with zero manual intervention; the model classifies reusable preferences into the semantic layer (covering categories such as habit, writing style, business requirements, and facts) and commits an episodic summary at the end of each task. Recall uses CJK-aware tokenization, importance weighting, and a 45-day exponential decay score, handling both single-character and bigram Chinese words as well as English tokens. Duplicates above a similarity threshold are merged automatically to keep the semantic layer lean. All data persists as plain JSON in your user directory - straightforward to back up and migrate, with no dependency on external services or vector databases. An HTTP API is also provided for browsing, searching, editing, and statistics.

It is built for heavy dsh users doing technical writing, business analysis, or multi-turn collaboration who want the agent to retain personal preferences and reduce repetitive context-setting. Zero configuration, zero build step, pure JavaScript - install and go, uninstall and clean up.