dsh-local-memory

Every new DSH session starts from scratch, forcing users to re-paste context and re-introduce themselves over and over. dsh-local-memory captures each turn as a concise markdown bullet, persists it to a local date-stamped file, and auto-injects a snapshot of the last few days into the next session, so the model still recognizes you after a restart.

The workflow has three parts: capture summarizes the user-assistant exchange at every turn/end and appends it to the day file; inject feeds the last N days (default 2, capped at 40 lines per file) into the system context as the very first step of a new session; memory_search registers as a tool the model can call on demand, scanning all date files with pure-JS keyword and substring matching and returning scored, ranked hits, making it easy to retrieve memories outside the injection window. The directory layout and anchor format are aligned with the opencode memsearch plugin, so the two can coexist, read from each other, and never overwrite.

Built for DSH users who run ongoing development, research, or daily assistant workflows and need real cross-session continuity. The entire plugin is pure JavaScript and plain text with zero native dependencies, requiring no pymilvus, numpy, or x86-64-v2 ISA baseline, so it runs out of the box on older CPUs and ARM.