dsh-tool-memory

Every DeepSeek Harness session starts amnesiac. The project conventions, user preferences, and workflow details you carefully shaped in one window are gone the moment you open the next. dsh-tool-memory ports the Hermes Agent memory mechanism into DSH, storing bounded plain-text entries (separator-delimited lines across two files: MEMORY for agent notes, USER for the user profile, capped at 2200 and 1375 characters respectively) so the agent genuinely remembers across sessions. The key design choice is the frozen snapshot: at session start the memory block is baked into the system prompt, while mid-session writes hit disk immediately without touching that prompt. This keeps the LLM prefix cache warm so you are not paying token re-encoding costs every single turn, yet the agent still sees up-to-date facts on its next call.

In practice the plugin registers six natural-language tools (add, replace-by-substring, remove, atomic batch, show, refresh) plus a first-use onboarding flow: when USER.md is empty, a guided section in the system prompt prompts the model to ask the user for name, role, and communication preferences; once the profile has content that section vanishes automatically, no file cleanup needed. A background consolidation loop adds a layer of autonomy: every N user messages (default 10) an isolated sub-agent, whitelisted to memory_* tools only, scans the last 24 turns for durable facts and preferences and merges them into the store. Writes are guarded by in-process promise mutex, atomic rename, drift protection (manual edits are respected, never silently overwritten, a .bak snapshot is kept), and a write-time threat scan that rejects injection, exfiltration, or hard-coded-secret patterns before they land in the file.

This plugin is aimed at anyone running extended development sessions, conversational workflows, or agent orchestration in DeepSeek Harness who is tired of re-explaining tech-stack conventions, documentation workflows, code-style preferences, or communication norms at the top of every new chat. Those scattered details are now captured automatically, persist across sessions, and remain inspectable as a plain-text file you can open in any editor any time.