I built Lore because my AI agents kept making the same mistakes. Agent A discovers that Stripe rate-limits at 100 req/min and figures out the backoff strategy. Next day, Agent B hits the same wall. No learning transfer.

Lore is a small library (not a service) that gives agents shared memory of operational lessons. One line to publish a lesson, one line to query it. PII is automatically redacted before storage.

Key design decisions: - Local-first — SQLite + ONNX embeddings, no server required. pip install lore-sdk and go. - Semantic search — Query by meaning, not keywords. "how to handle rate limits" finds the Stripe lesson. - Auto-redaction — API keys, emails, credit cards stripped automatically before storage. - Both languages — Python and TypeScript SDKs with cross-compatible SQLite DBs. - Optional server — Phase 2 adds a FastAPI server with PostgreSQL/pgvector for org-wide sharing across machines. The local SDK keeps working standalone.

What it's NOT: conversation memory (see Mem0/Zep), a vector database, or a RAG framework. It's specifically for structured operational lessons — "what went wrong and how we fixed it."

The SDK is ~500 lines per language. 258 tests. MIT licensed.

Would love feedback on the API design and whether this is a real pain point for others building with agents.