cookbook/08_learning/11_composition/TEST_LOG.md
Tested 2026-07-25 against gpt-5.5 (OpenAIResponses), branch feat/entity-memory-revamp, Postgres (pgvector container on 5532). Re-tested 2026-07-26 after the missing-model fix.
Status: PASS
Result: With no learning=, the hand-placed tools captured the preference (user memory) and the Meridian project + Priya link (entity memory); the printed manual-door surfaces show the guidance block and a data block whose relevance recall expanded Meridian for the message "what about meridian?" with the one-hop "runs <- Priya" edge.
2026-07-26 correction: the first log was wrong about the user memory. The machine
carried no model=, so update_user_memory returned "No model provided for memories
extraction" and stored nothing - only the entity write (no model needed) landed. The
same hole always_capture.py hit below; the manual door injects nothing. Re-run with
model= on the machine: update_user_memory stored "Prefers sources with primary data",
verified in the learnings table.
Status: PASS
Result: One deliberate order: learning tools + fs tools + both instruction blocks. The agent wrote notes/vector-db-comparison.md with the deadline. Model behavior note: it also wrote the "conclusions first" preference INTO the note alongside saving it - the one-claim-one-home discipline is exactly what the second-brain instructions add on top.
2026-07-26 correction: same missing model=, so the note was written but the
preference was not stored. Re-run with the model: update_user_memory(task=User prefers conclusions first in every summary.) and append_file(notes/tasks.md) both fired, and
the memory is in the table.
Status: PASS
Result: build_context() placed via additional_context, no tools: the read-only agent answered from its own knowledge.
2026-07-26 correction: the earlier "despite the seeded memory in the context, the
summary put its conclusion last" reads a model failure into a store failure - the seeding
call needed a model too, so nothing was seeded and the context block was empty. With
model= on the machine the seed lands; the answer still leads with its list and closes
with the conclusion, so the original observation (a data-only block informs but does not
compel) holds on the re-run.
Status: PASS
Result: post_hooks=[learning.capture_hook()] ran ALWAYS extraction in the background: profile (Name/Preferred Name: Dana) and one memory (data engineer, Lisbon, ClickHouse pipelines) appeared without any tool call. First run FAILED with empty stores - the manual door injects nothing, so the machine needed model= passed explicitly; the file and README now say so.
2026-07-26: this was the only file that had learned the lesson. LearningMachine.get_tools()
now warns once when a store that captures through a model has none, so the next person
finds out at attach time instead of from an empty table.