memory-knowledge-maintenance

name	memory-knowledge-maintenance
description	Maintain Kazusa's curated shared/world/common-sense memory knowledge base. Use this skill whenever the user asks to add, collect, edit, search, remove, deduplicate, sync, upload, or inspect world knowledge, common sense, local slang, character-world facts, or durable curated entries for the MongoDB memory collection. Prefer the local JSONL source and maintenance CLI over ad hoc MongoDB edits.

Memory Knowledge Maintenance

Use this skill to maintain the curated shared knowledge that RAG2 reaches through Memory-search and the persistent-memory helper agents.

The design is intentionally simple:

personalities/knowledge/memory_seed.jsonl  -> source of truth
MongoDB memory collection    -> indexed runtime copy
RAG2 Memory-search           -> retrieval interface

Do not introduce extra categories or dispatcher prefixes for common sense, world facts, or domain notes. If the most relevant memory row is retrieved, the existing persistent-memory search path can use it.

Files And Commands

Source file:

personalities\knowledge\memory_seed.jsonl

Maintenance CLI:

venv\Scripts\python.exe -m scripts.manage_memory_knowledge <command>

Useful commands:

venv\Scripts\python.exe -m scripts.manage_memory_knowledge validate
venv\Scripts\python.exe -m scripts.manage_memory_knowledge list
venv\Scripts\python.exe -m scripts.manage_memory_knowledge search 洗车
venv\Scripts\python.exe -m scripts.manage_memory_knowledge add --name "UCCU含义" --content "UCCU 是中文网络语“你看看你”的谐音/缩写，常用于调侃对方当前状态或行为。"
venv\Scripts\python.exe -m scripts.manage_memory_knowledge remove --name "UCCU含义"
venv\Scripts\python.exe -m scripts.manage_memory_knowledge sync
venv\Scripts\python.exe -m scripts.manage_memory_knowledge sync --apply

Use sync without --apply first. It is a dry run. Use --apply only after the local JSONL validates and the planned counters look right.

To make MongoDB match the local curated global source and remove unmanaged global rows:

venv\Scripts\python.exe -m scripts.manage_memory_knowledge sync --apply --prune-unmanaged-global

Entry Format

Each JSONL line is one MongoDB memory row:

{"memory_name":"洗车难题","content":"洗车难题是一个隐含约束推理题：如果问题是在普通情境下问“去洗车店洗车是走过去还是开车去”，默认应开车去，因为洗车服务对象是车，车必须到店。步行只有在额外前提成立时才合理，例如店家提供上门取送、代驾或用户只是先去查看/预约。","source_global_user_id":"","memory_type":"fact","source_kind":"external_imported","confidence_note":"人工整理的中文世界知识，用于覆盖该题的常见错误回答。","status":"active","expiry_timestamp":null}

Rules:

• Keep source_global_user_id empty. This file is for global/shared knowledge, not user-scoped profile memory.
• Use memory_type="fact" by default.
• Use memory_type="defense_rule" only for durable behavioral/reliability guardrails.
• Use source_kind="seeded_manual" for hand-curated common sense and local knowledge.
• Use source_kind="external_imported" for entries summarized from external references.
• Keep one entry to one useful idea. Concise entries retrieve better.
• Use stable, human-readable memory_name values. The de-duplication key is (memory_name, source_global_user_id).

What To Add

Prioritize high-leverage entries:

• Observed weak-LLM failures and their corrected rule.
• Common-sense hidden constraints and everyday causality.
• Local slang, memes, abbreviations, and group-specific terms.
• Stable character-world facts not better served by user_profile_agent.
• Recurring technical/domain notes from the community.
• Reliability guardrails such as "current prices/laws/version facts require live search."

Avoid:

• User-specific facts, preferences, health, impressions, commitments, or relationship information. Those belong in user_profile_memories.
• Volatile facts unless expiry_timestamp is set.
• Long articles or broad encyclopedia dumps.
• Duplicate paraphrases of the same idea.

Workflow

1. Search before adding:

   venv\Scripts\python.exe -m scripts.manage_memory_knowledge search <term>
   

2. Add or replace the local row:

   venv\Scripts\python.exe -m scripts.manage_memory_knowledge add --name "<name>" --content "<content>"
   

3. Validate:

   venv\Scripts\python.exe -m scripts.manage_memory_knowledge validate
   

4. Dry-run sync:

   venv\Scripts\python.exe -m scripts.manage_memory_knowledge sync
   

5. Apply sync:

   venv\Scripts\python.exe -m scripts.manage_memory_knowledge sync --apply
   

When the user asks to make the database match the local curated file, use:

venv\Scripts\python.exe -m scripts.manage_memory_knowledge sync --apply --prune-unmanaged-global

Before destructive DB cleanup, export a backup with:

venv\Scripts\python.exe -m scripts.export_memory --limit 1000 --output test_artifacts/memory_before_<short_task_name>.json