The agent keeps the approved store on the first attempt.
This is the guided-first-pass variant. The project runs on edge hardware and keeps state in human-editable JSON. Mneme supplies ADR-001 before generation, so the agent extends the module already there instead of reaching for a database and migration stack.
The risky architecture never gets written.
Guidance changes the first governed output. The task stays inside the approved boundary and passes the check.
ADR-001 · JSON storage only. No external database on edge hardware. ADR-003 forbids an ORM in v1; ADR-005 requires extending the current storage module before rebuilding it.
-
01 Decision
The repository records JSON as an operational constraint, not an accidental implementation detail.
ADR-001 · JSON only · edge deployment -
02 Guidance
Mneme selects the three applicable decisions before the agent plans the refactor.
ADR-001 + ADR-003 + ADR-005 · extend the current storeDemonstrated -
03 Violation
Without project guidance, the training-data prior is a database migration.
Counterfactual: PostgreSQL + Redis + ORM + migrationsIllustrative comparison -
04 Deny / warnN/A
No violating governed write is evaluated in this walkthrough; guidance prevents it earlier.
Not present in this scenario -
05 CorrectionN/A
No retry is required because the first governed plan already respects the decision.
Not required -
06 AllowPASS
The agent extends the JSON module and introduces no external persistence dependency.
Result: PASS · storage_violation_count=0
Run it yourself
Verify the compliant storage task
The product repository includes the decision corpus and clean storage prompt used by the flagship trace.
git clone https://github.com/MnemeHQ/mneme
cd mneme
pip install -e .
mneme check --memory examples/project_memory.json \
--input examples/prompt_clean.txt --query "storage backend"
# Result: PASS ยท exit 0This supporting page proves the aligned output. The flagship storage demo runs both the violating and corrected inputs.
What happened
The model does not need to guess whether JSON is temporary; the decision is explicit before generation.
The counterfactual migration explains the risk, but this page does not pretend a DENY happened.
No database service, credential path, migration history, or second source of truth is introduced.
Common questions
Why is there no DENY stage?
Because this example demonstrates pre-generation guidance. The governed agent proposes a compliant JSON extension on its first attempt, so there is no violating write to deny.
What does the PASS prove?
It proves that the resulting task contains no forbidden database, ORM, or migration pattern under the recorded storage decisions.
Can the deny-to-pass sequence be reproduced?
Yes. The governed Python agent flagship runs both the violating and corrected fixtures through the same decision.
Check the decision before the refactor.
Record the storage boundary once, then use the same corpus in agent guidance and CI.
pipx install "mneme-hq>=0.5.1"