cookbook(openai-agents): use gpt-4.1-mini + sharpen Pattern-1 prompts

[DK09876]

Summary

Two related changes to notebooks/10-openai-agents-memory.ipynb to make the memory demos work reliably end-to-end. Both pair with the integration PR vectorize-io/hindsight#1866.

What & why

1. Model upgrade: gpt-4o-mini → gpt-4.1-mini (cells 4 and 9)

Pattern 2 (memory_instructions auto-inject) was returning "I currently don't have that information" under gpt-4o-mini even though memory_instructions had injected the recalled facts into the system prompt — the model was simply not grounding its answer on them. Under gpt-4.1-mini, cell 10 now returns the recalled facts verbatim:

"You use Python and SQL as your programming languages. As for tools, you use VS Code as your code editor, and you prefer to work in dark mode."

Pattern 2 is the load-bearing demo for the auto-inject path most users will actually adopt.

2. Pattern-1 prompt sharpening (cells 7 and 8)

Same template the Claude SDK cookbook adopted (commit d92ccab373 on cookbook main): nudge the agent to call recall_memory explicitly before answering, instead of leaving the tool-use decision to the model.

• Cell 7: "What IDE do I use? And what's my job?" → "Recall what you know about me first, then answer: what IDE do I use and what's my job?"
• Cell 8: same template ("Use the recall tool to refresh what you know about me first, then recommend...").

Cell 7 stays best-effort under gpt-4.1-mini's tool-use variance — the model still sometimes asks "what's your IDE?" rather than recalling — but with the sharpened prompt the failure rate drops and the cell makes the user's intent explicit to a reader.

Relationship to other PRs

• Pairs with fix(openai-agents): default to Cloud + gated E2E + requires_real_llm bucketing hindsight#1866 (the integration code is unchanged; only the cookbook is touched here).
• Supersedes the unmerged PR cookbook: sharpen Pattern-1 prompts in 08-llamaindex + 10-openai-agents #38 (cookbook/pattern1-prompt-polish-oa-llamaindex) for the OpenAI Agents notebook. That PR was abandoned because the prompt polish alone didn't reliably move Pattern 1 cells under gpt-4o-mini. This PR pairs the polish with the model bump, which is what unlocks Pattern 2 reliably.

Test plan

• Cookbook 11/11 cells execute against canonical PR #1866 build, 0 errors.
• Cell 10 (Pattern 2 auto-inject) reliably returns recalled facts (Python/SQL/VS Code/dark mode).
• Cell 7 (Pattern 1 tool recall) is best-effort under gpt-4.1-mini; cookbook narrative now matches what the user can expect.
• Reviewer to run end-to-end against the merged hindsight-openai-agents.

🤖 Generated with Claude Code