Add Agent Memory context and memory engineering lab (OAMP 26.8 + Ollama) - #2

Summary

Add Agent Memory: Introduction to context and memory engineering, a guided Oracle LiveLabs workshop in ingestion/jl_notebooks/agent_memory_zero_to_hero.

• Provide matching student and complete Ollama notebooks, organized into nine Parts with 17 numbered TODOs, a learning roadmap, expandable hints and explanatory Markdown before every code cell.
• Use an understandable ecommerce inventory handover: 80 mugs proposed, 120 approved, and no supplier order submitted. Build progressively from a bare LLM to a transient tool loop and then a memory aware agent.
• Demonstrate OAMP 26.8 with LangChain, Ollama/Llama 3.2 and in-database embeddings/vector retrieval. Cover working, episodic, semantic and procedural memory; extraction, reviewed supersession, linked-history recall, offloading, compaction, bounded retrieval, pruning, context cards, observability and retention.
• Include the synthetic fixture, seven editable/embedded architecture diagrams, pinned requirements, delivery notes and opt-in instructor provisioning. Remove the earlier reranker and Deep Data Security dependencies from this workshop.
• Use the same llama3.2 model through native chat and the compatible API, without Qwen-specific thinking controls. Include the structured-request check and repeat-safe procedure seeding for sequential cell reruns. Keep optional writes behind host-controlled approval of the exact handover draft.
• Check background extraction separately for each note using SDK error logs, unsuccessful operation events and fresh capture IDs, so older memories cannot mask a failed update. Clarify that context-card token budgets limit formatted relevant records, not the whole prompt.

Validation

• Both notebook schemas validate, and all code cells compile with notebook-compatible top-level await.
• Shared Markdown and supplied code match between editions; student outputs are empty; TODOs 1–17 and their hint dropdowns are unique and complete.
• Embedded SVGs match their companion files; SVG/fixture parsing and local README links pass.
• Shell syntax checks pass. The provisioning script's disabled/default path was exercised without database writes.
• git diff --cached --check and a credential-pattern scan of all 17 publishable files pass. No .env, credentials or local virtual environments are included.

Execution evidence: a fresh-kernel Llama 3.2 run on 9 October 2026 completed 80/80 code cells without cell exceptions (run 3738399430, about 91 seconds excluding the model download). Actual outputs are saved in the existing complete notebook; student outputs remain empty. The run used OAMP 26.8.0, LiteLLM 1.104.0, Ollama 0.40.0 and local Oracle AI Database with in-database embeddings.

Same core inventory answer, not equivalent overall quality. Compared with the previous Qwen3.5 9B run (22d0a021ed, preserved at commit 1cac500), the prompts, sampling settings, fixture and database workflow were unchanged:

• Memory-aware initial and fresh-session answers retained 120 units approved under PO-1048, no supplier order submitted. Procedure Hit@3 was 1 for all three questions, and scoped cleanup left zero rows for both case threads.
• Both per-ingestion checkpoints passed with two candidates each. However, the Llama extraction omitted the original proposal quantity and explicit no-order statement; its compact summary also omitted that explicit negative.
• The no-memory agent invented 12 units and an approval reference of SOP-102. The memory-aware trace contains one handled tool-validation error for an invalid inspect_stock_report argument before it recovered the correct answer.
• The final handover preserved the business facts but altered one source UUID and exceeded the under-120-word instruction (129 whitespace-separated words). The raw output is kept visible; both editions now explain how to inspect error observations, omissions and citations.
• Focused ingestion-guard checks passed again. The earlier real SDK provider-failure and repeat-safe seed checks remain documented as historical checks, not new Llama tests.
• The unattended run declined the optional handover save; its approved-write path remains untested. A macOS process-inspection warning occurred during kernel shutdown after all cells completed, without preventing output saving.

The README includes the side-by-side observations and model digest. One run per model is not a reliability benchmark.

Delivery and boundaries

The hosted runtime must supply the pinned Python packages, a dedicated Oracle workshop account, the imported embedding model and the configured Ollama service. Existing image behavior is unchanged unless the new provisioning script is explicitly enabled.

User/thread labels demonstrate application scoping, not multi-user authorization. Stored memories and fluent model answers are not automatically verified facts. The agent does not submit supplier orders, and scoped cleanup does not prove erasure from backups or storage media.

The separate developer/dev-rel-con-mem-engineering workshop pages are not part of this PR.