Agent Memory Health Audit

Inputs

• - Memory JSONL following the common field contract.
• Policy JSON with lifecycle, similarity, quality, and redaction thresholds.
• Optional previous-state JSON containing still-open issue keys.
• A fixed --as-of date for reproducible age calculations.

Outputs

• - summary.json: issue counts and unchanged-input confirmation.
• INLINECODE2 and CSV: evidence, severity, and suggested human action.
• INLINECODE3: duplicate candidates only.
• INLINECODE4: conflicting facts requiring source verification.
• INLINECODE5, reverify_queue.json, and quarantine_candidates.json.
• INLINECODE8 and next_state.json.

Three-step workflow

1. 1. Validate the export and compute deterministic quality and lifecycle signals.
2. Review duplicate, conflict, archive, re-verification, and privacy queues.
3. Execute approved changes only through the authorized memory system, then export and audit again.

Run

CODEBLOCK0

Boundaries

The script does not call a memory API, external AI model, vector database, MCP server, or deletion endpoint. It never modifies the input export. Findings are conservative review signals, not automatic deletion decisions or legal/privacy conclusions.

See README.md and references/ for the schema, decision policy, recovery steps, and acceptance checks.