The patterns in this workflow are synthesized from multiple open-source implementations and community practices.
- Repository: seedprod/openclaw-prompts-and-skills
- Key contribution: The SOUL.md / USER.md / MEMORY.md separation, heartbeat protocol, memory security model (MEMORY.md main-session only), daily memory files with
YYYY-MM-DD.mdformat - Related: OpenClaw Heartbeat docs
| Repository | Focus |
|---|---|
| anthropics/skills | Official Anthropic skills |
| hesreallyhim/awesome-claude-code | Curated skills, hooks, and tools |
| VoltAgent/awesome-agent-skills | 500+ community skills |
| rsmdt/the-startup | Spec-driven development framework |
| trailofbits/skills | Security research skills |
| coleam00/excalidraw-diagram-skill | Visual diagramming |
| Resource | Description |
|---|---|
| Agent Skills spec | Open standard for agent skills (works across Claude, Codex, etc.) |
| ACOS v6 | Creator-focused OS with 630+ skills |
| OpenClaw workspace skill | AGENTS.md / SOUL.md / TOOLS.md / MEMORY.md maintenance |
| Article | Author | Key Insight |
|---|---|---|
| Complete Guide to AI Agent Memory Files | HackerNoon | Comparison of CLAUDE.md, AGENTS.md, and other memory file standards |
| Memory Architecture for Agentic Systems | spikelab | Comprehensive survey: Tulving's framework, MCP servers, cost analysis |
| OpenClaw 50-day workflows | velvet-shark | 20 real production workflows with prompts |
| Claude Code Best Practices | Anthropic | Official guidance on CLAUDE.md, skills, and agentic patterns |
| Agentic Heartbeat Pattern | Marcilio Mendonca | Hierarchical agent coordination (expansion/contraction cycle) |
| Extend Claude with Skills | Anthropic | Official skills documentation |
| Discover Plugins | Anthropic | Official marketplace documentation |
| Paper/Concept | Relevance |
|---|---|
| Tulving's Memory Taxonomy (1972) | Episodic / Semantic / Procedural → daily files / MEMORY.md / skills |
| Generative Agents (Park et al., 2023) | Observation → Reflection → Planning loop |
| Reflexion (Shinn et al., 2023) | Verbal reinforcement learning through self-reflection |
| Plain filesystem benchmark (Letta) | 74% on LoCoMo — simple markdown often beats complex systems |
"Plain filesystem achieves 74% on LoCoMo benchmark, beating specialized memory libraries." — Letta research
This workflow deliberately uses plain markdown files over databases, vector stores, or custom infrastructure. The benefits:
- Human-readable and inspectable
- Git-friendly (diffs, history, backup, collaboration)
- Zero infrastructure overhead
- LLM-native (models are trained on markdown)
- Portable across tools (works with Claude Code, Codex, Cursor, etc.)