memory-system

Author: yunyu422 · Built for DeepSeek Harness (DSH) and any host that supports SKILL.md skills

中文 | English

Bring WorkBuddy-style three-layer memory to DeepSeek Harness: user profile (permanent) → workspace memory (per-project) → auto-learned profile. Agents forget everything between sessions; this skill lets context truly accumulate across sessions and projects.

📡 Cloud sync (M1 shipped): the dsh-memory-sync plugin is available — Git backend + conflict-aware merge + optional E2E encryption, with memory_sync_pull/push folded into the session routines. See the Cloud Sync Backend Design (M2 WebDAV/S3 + encryption → M3 team sharing).

The three layers

Install (DSH)

# Copy the skill into the user skill root (applies to all projects)
Copy-Item -Recurse .\skill\memory-system "$env:USERPROFILE\.dsh\skills\memory-system"

You can also copy it into a single project's <projectRoot>/.dsh/skills/ to scope it to that project. DSH hot-refreshes the skill catalog on every request step — no restart needed; the memory-system entry appears in the skill catalog on the next session.

Usage

• Automatic: at session start the agent reads all three layers and applies preferences and conventions; at task end it writes the daily log and settles durable conventions automatically.
• Manual: just say —
  • "remember <fact>" / "what did we do today" / "project conventions" / "what are my preferences" / "forget <entry>" / "memory status"

Directory structure

memory-system/
├── SKILL.md                 # Main instructions: session start/end routines, write triggers, invariants
├── references/              # Per-layer detailed rules (loaded on demand)
│   ├── user-profile.md      #   L1 structure and maintenance rules
│   ├── workspace-memory.md  #   L2 formats, rotation, pruning
│   └── learned-preferences.md # L3 evidence/confidence/promotion/decay and optional cloud sync
└── templates/               # Templates for creating new files
    ├── USER.md
    ├── MEMORY.md
    ├── PREFERENCES.md
    └── daily-log.md

Design highlights

• Write triggers, not write-every-turn: L1 is written only when the user explicitly states a durable personal fact; L2 is written in the task-end routine; L3 uses the "explicitly stated or ≥2 observations" evidence rule so one-off choices are never misrecorded as long-term preferences.
• Privacy & security built in: no credentials/keys in any memory file; L1 is never rewritten wholesale without confirmation; "forget X" removes specific entries precisely.
• Optional cloud sync: a pure local skill cannot run a real cloud backend, but pointing ${DSH_HOME}/memory/ at a private Git repo or a synced folder gives cross-machine "cloud profile" behavior (see references/learned-preferences.md).
• Path rules match DSH: DSH_HOME defaults to ~/.dsh; project root is the nearest .git ancestor, falling back to cwd.

Compatibility

• The skill is a standard SKILL.md + frontmatter (name / description / user-invocable: true), matching DSH dsh-skill-filesystem discovery (<root>/<name>/SKILL.md, kebab-case names).
• Memory files are plain Markdown, directly readable by any toolchain (Claude Code / Codex / Cursor) for easy migration.

License

MIT