Profile Memory Ledger

Create a private, reviewable ledger of facts a user explicitly wants considered for future conversations. Run the local curator to verify evidence and classify proposed changes. Read the input format when preparing candidates.

Ask what material the user authorizes for this task and where the private output should go. Use only material provided or accessible with authorization. Do not put personal material into a public repository, shared log, or external model request.
Extract one atomic candidate per fact. Copy a short exact excerpt from the source into sources[].excerpt; put that same exact span in the candidate's evidence. Give each source a stable ID, a local or user-approved reference, and the observation date. Do not infer facts that the excerpt does not support. Skip secrets, credentials, financial account numbers, government IDs, and third-party private data.
Choose a policy for each candidate: stable (review in 365 days), periodic (review in 30 days), or temporary (review in 7 days). Treat these as prompts for review, not proof the fact is still true.
Write the candidate JSON to a private local path. Run python3 scripts/curate.py INPUT.json -o LEDGER.json. Optionally add --existing PREVIOUS.json to compare with a previous ledger. The script has no network access and uses only the Python standard library.
Read LEDGER.json. Show the user each add, refresh, and review proposal with its evidence and review date. A keep result is unchanged. Do not automatically save proposals to a persistent memory system. Save only facts the user approves, through the host's authorized memory mechanism. For conflicts, keep the prior value until the user decides; never silently replace it.
If the source material changes, rerun with a new observation date and --existing. If a review date has passed, verify the fact from a fresh authorized source before using it as current.

The script checks exact evidence containment and key conflicts. It does not determine whether the excerpt semantically proves a fact or whether personal data is safe to store; inspect those judgments before proposing or saving anything.