dsh-plugin-memory

In a multi-turn agent session, the context window resets after every turn, and the reasoning, tool calls, and lessons from the previous exchange are effectively lost. dsh-plugin-memory solves this by automatically distilling each completed turn: it collects the human question, the final assistant answer, and the tool calls, then uses an auxiliary model call to compress them into a structured knowledge entry with a title, answer, and sources, persisted as a Markdown file in a memory directory.

What sets it apart is its continuous accumulation. When the auxiliary model judges a freshly distilled entry to be on the same topic as an existing one, a second fusion call merges the two into a single, richer record; if the fusion fails, the new entry simply supersedes the old one, so no data is lost. Index writes are atomic (temp file plus rename), and entry writes are serialized through a single promise chain for concurrency safety. When the ctx.skills service is mounted, the plugin also registers each memory entry as a memory- skill candidate, allowing the model to browse its durable knowledge on demand.

Built for developers using DeepSeek Harness to power multi-turn agent applications who want domain knowledge to survive across sessions without manually managing memory files. A turn ends, and the insight is already captured.