dsh-memory-plugin

Every Harness session ends with its context evaporating. Project conventions, user preferences, and the conclusions reasoned out in the previous turn all have to be re-explained from scratch. dsh-memory-plugin turns that gap into a durable, recallable, and tunable memory layer: a manually curated zone for project facts and user preferences, an auto-retention zone that distills completed conversations into dynamic facts, an inbox prefetch that stages relevant context before the current question even ships, and a four-channel weighted recall pipeline combining semantic, lexical, exact-entity, and graph signals to inject the most pertinent fragments into the model request within a controllable character budget.

The plugin writes to an isolated ysh_memory storage domain so it never clobbers records owned by other memory plugins. An optional MemoryEmbedder can be registered externally to layer vector scoring on top without replacing the local graph, and external providers can swap in custom retention or recall implementations through a public registry, keeping the architecture genuinely pluggable.

It fits engineers who need project continuity across many sessions, product teams that want to sediment user preferences into lasting agent context, and self-hosting operators who prefer local long-term memory without adding an external SaaS dependency. As long as the Profile already ships the standard storage and LLM services from dsh-base, the plugin is ready the moment it is installed.