I've been working on Skill Bank - an open-source platform for building AI agents with dynamic capability discovery and user memory.
*The Problem:*
Most AI agent frameworks hardcode tool lists and configurations. Every time an agent executes a task, users must provide full parameter sets. There's no learning, no personalization, no memory.
*What Skill Bank Does:*
1. *Semantic Discovery*: Agents find capabilities via embeddings + RAG, not hardcoded lists.
2. *The "Golden Rule"*: Instead of creating dozens of near-identical tools (create_user, read_user, etc.), use 1 atomic tool + N specific skills. This preserves semantic diversity and improves RAG.
3. *Memory & Learning* (v1.5): System automatically learns user patterns: - Detects preferences after 5 consistent executions - Auto-fills missing parameters (70%+ confidence) - Per-user preference profiles - Result: 60% fewer inputs after learning
4. *RAG Integration*: Context-aware skills query indexed documents automatically.
5. *Production-Ready*: 128 tests (100% critical passing), quality gates enforced, complete docs.
*Architecture:*
6 layers (4 implemented): - Tools (atomic capabilities) - Skills (workflows/knowledge) - Documents (RAG) - Memory & Learning - (Coming: Credentials, Sub-Agents)
*Tech:* TypeScript, SQLite + sqlite-vec, Vitest, ~12K LOC, MIT License.
*Quick Start:* ```bash git clone https://github.com/MauricioPerera/Skill-Bank npm install npm run demo:memory # See learning in action ```
*Demo Output:* ``` After 5 executions - Pattern detected! Learned 3 preferences for Alice: • format: "PDF" (confidence: 100%) AUTO-FILLED PARAMETERS: • format: "PDF" (100% confident) ```
Would love feedback from the HN community! What would you add to an agentic platform?