dsh-memory-manager

Multi-turn conversations in DeepSeek Harness Web UI tend to lose user preferences, project conventions, and past feedback once a session ends. dsh-memory-manager adds a persistent, per-project memory layer to DSH: each entry is stored as a Markdown file with a type tag (user, feedback, reference, project, other) and a project scope, surviving across sessions so the agent always knows what matters.

Core capabilities include a dedicated memory management page in Settings with search, project and type filters, and full create-edit-delete workflow; automatic index injection of the current project and global memories into the system prompt on every model call, so the agent only fetches full text on demand, saving roughly an order of magnitude of context tokens compared to stuffing every entry verbatim; three registered agent tools (memory_save, memory_get, memory_list) that let the model proactively capture user preferences and project knowledge during conversation; and a one-click import that scans the ZCode CLI memory directory, deduplicates by name, and brings accumulated history into DSH with minimal friction.

It is well suited for DSH users who need consistent context across sessions, and especially for those migrating from the ZCode ecosystem who want to carry over accumulated user preferences and technical references into a new workflow without re-typing everything.