dsh-hermes-memory

Every new session in DeepSeek Harness starts from a blank slate. The pitfalls you hit during the last debugging run, the coding style the user prefers, project-specific naming conventions—all of it evaporates the moment a conversation ends. dsh-hermes-memory fills that cross-session gap: it distills high-value context into structured, persistent memories so the agent walks into the next conversation already carrying its full history.

The plugin ships four explicit tools—memory_save, memory_search, memory_replace, and memory_remove—across four scopes: global, user, project, and failure. Every record is written into the DSH-native storage domain and validated against a schema before commit, with a pre-write secret and prompt-injection scan catching sensitive or malicious content before it ever hits disk. From V3 onward, a bounded reference context can be injected at session start; V4 adds rule-based automatic capture that recognises corrections, project conventions, and preferences from real message events and archives them alongside provenance metadata such as session ID and event sequence; and V4.2 introduces retention-driven cleanup anchored on lastReferencedAt, so the memory store keeps what matters and lets the rest go.

If you use DeepSeek Harness for daily development and want the agent to retain project norms and reduce repetitive corrections, or your team needs an auditable, provenance-tracked memory layer, this plugin provides a lightweight, ready-to-use foundation. It does not replace the chat UI, does not introduce vector or embedding retrieval, and stays lean by integrating deeply with the DSH storage domain—making it a practical, quiet long-term memory layer for your workflow.