{"skill":{"slug":"elite-human-memory","displayName":"elite-human-memory","summary":"Portable Clawhub-compatible human-like memory system with layered storage (Working/Episodic/Semantic + optional Vector index), rich metadata schema, auto-pro...","description":"---\nname: elite-human-memory\ndescription: Portable Clawhub-compatible human-like memory system with layered storage (Working/Episodic/Semantic + optional Vector index), rich metadata schema, auto-promotion heuristics, conflict detection & resolution, and semantic retrieval. Platform-agnostic design suitable for publishing and multi-agent use.\n---\n\n# Elite Human Memory — Portable Edition\n\nThis version prioritizes broad compatibility and portability. It preserves the human-like, selective, and contextual philosophy while avoiding tight coupling to any single agent framework. Ideal for Clawhub, custom agents, or cross-platform deployments.\n\n## Memory Layers\n\n**Working Memory**  \nCurrent conversation only. Transient and not persisted beyond the session.\n\n**Episodic Memory**  \nDaily raw memory files stored at:  \n`memory/daily/YYYY-MM-DD.md` (or user-configurable path)\n\n**Semantic Memory**  \nCurated long-term memory stored in:  \n`memory/MEMORY.md`\n\n**Vector Index** (Optional / Portable)  \nEmbeddings generated from `MEMORY.md` and recent episodic files.  \nCan be implemented with any vector database (Chroma, FAISS, LanceDB, or even a simple embedding cache).  \nStored in: `memory/vectors/` or external vector store.\n\nThis enables semantic search alongside traditional metadata filtering. If no vector capability is available, fall back to keyword + metadata search.\n\n**Conflict Ledger**  \nDetected contradictions are logged in:  \n`memory/conflicts/`\n\n## Context Schema (Metadata)\n\nEvery memory entry should include:\n\n- **When**: Timestamp + recency weight\n- **Where**: Channel/context (Telegram, CLI, web, Discord, etc.)\n- **Why**: Trigger or reason it was recorded\n- **State**: `active` | `stale` | `superseded` | `resolved`\n- **Scope**: `global` | `project` | `person` | `temporary`\n- **Validity**: \n  - `confidence`: high / medium / low\n  - `last_verified`: date\n  - `expires`: optional date\n- **Related**: Links to other memories, people, or projects\n- **Source**: Path + line number or reference (for traceability)\n\n## Auto-Promotion Heuristics\n\nThe agent should evaluate daily memory entries for promotion using the following weighted signals:\n\n**Strong signals (high weight):**\n- Explicitly referenced by the user in later conversations\n- Repeated across 3+ days or sessions\n- Tied to a core project, goal, or person\n- User corrects or reinforces the memory\n\n**Supporting signals (medium weight):**\n- High confidence rating\n- Clear future utility\n- Related to an active decision or preference\n\n**Promotion Rules:**\n- If 2+ strong signals → Propose promotion\n- If 1 strong + 2 supporting signals → Propose promotion\n- If only supporting signals → Log for weekly review only\n\n**User Control:**\n- Default behavior: Always propose before promoting\n- Optional mode: `auto_promote = true` (for trusted, low-risk memories)\n\n## Conflict Detection & Resolution\n\nWhen two memories contain contradictory information, the agent should:\n\n1. **Detect** the conflict during write or weekly maintenance.\n2. **Log** it in `memory/conflicts/` with:\n   - Both conflicting statements\n   - Context and sources\n   - Severity (high / medium / low)\n3. **Resolve** using one of these methods:\n   - Ask the user for clarification\n   - Propose a resolution with reasoning\n   - Auto-resolve low-severity conflicts with a note (e.g. “Temporarily preferred X over Y”)\n\nAll resolutions must update the `State` field of the affected memories and record the decision in the conflict log.\n\n## Retrieval Strategy\n\nWhen the user asks about history, decisions, preferences, or past context, the agent should follow this order:\n\n1. **Semantic Search** (Primary)  \n   Query the vector index (or embedding similarity) over `MEMORY.md` and recent daily files for relevant memories.\n\n2. **Metadata Filtering** (Secondary)  \n   Apply filters on scope, state, confidence, date, and related entities.\n\n3. **Keyword / Full-text Fallback**  \n   Traditional search when vector capabilities are unavailable.\n\n4. **Response Guidelines**  \n   - Use natural confidence language  \n   - Mention if a memory may be stale or conflicting  \n   - Include `Source:` references when helpful\n\n## Platform Integration Notes\n\nThis skill is designed to be **framework-agnostic**:\n\n- Use file system operations for storage and retrieval\n- Integrate with any available embedding model or vector store for semantic search\n- Can coexist with simpler key-value memory systems (map rich contextual memory here, simple facts elsewhere)\n- Works in Hermes, Clawhub-published agents, LangChain, AutoGen, or custom loops\n\n## Behavioral Triggers\n\n**Auto-read memory when:**\n- User asks about past decisions, preferences, people, projects, or dates\n- The current context feels incomplete or contradictory\n\n**Auto-write memory when:**\n- User gives explicit “remember this” instructions\n- Clear decisions or repeated preferences appear\n- New long-running context is established\n\n**Auto-maintenance:**\n- Weekly review (can be manually triggered or scheduled via cron/agent loop)\n\n## Storage Layout (Example)\n\n```\nmemory/\n├── daily/\n│   └── YYYY-MM-DD.md              # Episodic / daily memory\n├── conflicts/\n│   └── YYYY-MM-DD.md              # Conflict logs and resolutions\n├── MEMORY.md                      # Curated long-term semantic memory\n└── vectors/                       # Optional embeddings (or use external DB)\n```\n\nPaths are examples — make them configurable for the target environment.\n\n## Notes\n\n- This skill is optimized for portability and easy publishing on Clawhub.\n- Vector search (when available) significantly improves retrieval quality.\n- Auto-promotion and conflict detection reduce manual maintenance burden while keeping the user in control.\n- No hard dependencies on proprietary tools or specific runtimes.\n\nSee `references/two-track-approach.md` for the parallel Hermes-optimized vs portable versioning strategy used in this skill family.\n\n---\n\n**Version:** 1.0.0 (Portable / Clawhub Compatible)  \n**Status:** Ready for publishing and cross-platform use","tags":{"latest":"1.0.1"},"stats":{"comments":0,"downloads":839,"installsAllTime":2,"installsCurrent":1,"stars":0,"versions":2},"createdAt":1773102519480,"updatedAt":1779166126094},"latestVersion":{"version":"1.0.1","createdAt":1779166115529,"changelog":"- Portable, Clawhub-compatible version released; prioritizes framework-agnostic, cross-platform use\n- Supports layered storage: Working, Episodic, Semantic memory, and optional vector index for semantic search\n- Adds rich metadata schema, auto-promotion heuristics, and conflict detection/resolution workflow\n- Retrieval strategy now prioritizes semantic search, with metadata filtering and keyword fallback\n- Updated maintenance, behavioral triggers, and integration notes for broader agent and environment compatibility\n- Updated documentation and storage layout; see README.md and INTEGRATION.md for details","license":"MIT-0"},"metadata":null,"owner":{"handle":"cliftonwknox","userId":"s173hc1n2mersc48zpy1pq1czn83m308","displayName":"cliftonwknox","image":"https://avatars.githubusercontent.com/u/133707677?v=4"},"moderation":{"isSuspicious":false,"isMalwareBlocked":false,"verdict":"clean","reasonCodes":["review.llm_review"],"summary":"Review: review.llm_review","engineVersion":"v2.4.24","updatedAt":1780090838498}}