richmondalake/memorizz/memory-curator

Curate project memory

Work on the project's memory ID (from the session context, or memorizz plugin memory-id).

1. Inventory. memorizz_memory_status for counts, then

memorizz_list_memories (memory_type "knowledge_base", a high limit) to read every current fact.

1. Group facts about the same thing (a command, a service, a decision).

Within each group look for:

• duplicates: the same fact saved twice;
• contradictions: two facts that can't both be true;
• stale facts: ones the code now disagrees with. Check the code, config

or docs before calling a fact stale.

1. Fix, one change per call:

• duplicates: keep the best-worded record (if another has detail it

lacks, first memorizz_update_memory it with the merged text and keep the new record), then retire each other copy with memorizz_update_memory (record_id the copy, duplicate_of the kept record, no content);

• contradictions: memorizz_update_memory the wrong one with the right

fact, or retire it with duplicate_of the record that is right;

• stale facts: memorizz_update_memory with the current truth and the reason

("changed in commit …" or "per pyproject.toml");

• facts that should not exist at all (secrets, personal data, wrong project):

memorizz_forget_memory, which asks the user to approve.

1. Entities. Facts that describe a person, service or system with

attributes belong as entities: memorizz_upsert_entity, then supersede the loose fact.

1. Report what you merged, corrected and proposed for deletion, with record

IDs, and anything you weren't sure about. Never delete or supersede on a guess; ask.