Add notebook 12: Hindsight on Oracle 23ai

[DK09876]

What

A new cookbook notebook, notebooks/12-oracle-backend.ipynb, showing Hindsight running on Oracle Database 23ai — memory stored and searched natively in Oracle (a VECTOR column with an HNSW index, Oracle Text for keyword search, and the entity graph in ordinary tables — no separate vector database).

Written for an audience new to Hindsight, using a customer-support example (Acme Corp). It walks through each feature and, for this Oracle-focused audience, shows the raw SQL underneath each step:

• retain — teach the agent facts in plain English; LLM extraction into MEMORY_UNITS (native VECTOR) + ENTITIES.
• recall's lenses, each shown as native Oracle SQL:
  • semantic search — VECTOR_DISTANCE(... COSINE) + HNSW
  • keyword search — Oracle Text CONTAINS()
  • tag scoping, the entity knowledge graph, and temporal/recency (event_date)
• recall — one fused call (RRF + reranking) returning the memories relevant to a specific question.
• reflect — the payoff: it reasons across three separately-logged incidents and surfaces a root cause (lock contention on the ORDERS table during overnight jobs) that appears in no single memory, then recommends a fix.
• mental models — a durable, auto-refreshable summary stored in Oracle with its own VECTOR.

Notes

• The notebook talks to a Hindsight server configured with the Oracle backend (HINDSIGHT_API_DATABASE_BACKEND=oracle + a connection URL) and uses local sentence-transformers embeddings, so it's self-contained against a local Oracle 23ai.
• Dependency: the recall (linked-memory) and mental-model paths rely on Oracle backend fixes in hindsight-core (fix(oracle): make recall and mental-model history work on the Oracle backend hindsight#1980). Best merged after those ship.

🤖 Generated with Claude Code