mathclaw-ruc/mathclaw/memory

Memory

Structure

• memory/daily_memory/YYYY.M.D/YYYY_M_D.md - Human-readable daily summary for one day.
• memory/daily_memory/YYYY.M.D/YYYY_M_D.json - Structured daily payload for code and downstream features.
• memory/weekly_memory/YYYY_M_D_to_YYYY_M_D/YYYY_M_D_to_YYYY_M_D.md - Weekly study plan generated from the last 7 daily summaries.
• memory/graphs/knowledge_graph.json - Knowledge-point graph.
• memory/graphs/error_graph.json - Error-pattern graph.
• memory/MEMORY.md - Auto-generated compatibility snapshot loaded into prompt context.
• memory/HISTORY.md - Append-only audit log of consolidation events.

Read Order

Prefer these sources in order:

• Daily .json when you need structured fields such as date, high_risk_knowledge_points, high_frequency_error_types, learning_status_summary, and tomorrow_study_suggestions.
• Daily .md when you need a human-facing summary.
• Graph JSON files when you need relationships, ranking, or historical nodes.
• Weekly .md when you need the next-week learning plan.
• HISTORY.md only when you need an audit trail of what was archived.

Graph Semantics

Knowledge graph:

• Nodes are knowledge points.
• Important fields include time_points, risk, mastery, importance, last_seen, and display_size.
• Relations include prerequisite, similar, contains, and related links.

Error graph:

• Nodes are error patterns.
• Important fields include error_count, severity, repeated, last_seen, and display_size.
• Relations include corresponding knowledge point, similar error, and correction suggestions.

Dynamic Adjustment Rules

• Low-frequency, low-risk, long-inactive nodes can be archived.
• Similar concepts should be merged upstream before adding too many nodes.
• Example items only keep recent representative samples; older examples are dropped automatically.
• Frontend views should size nodes by importance or severity, not raw node count.

Update Policy

• Do not manually edit MEMORY.md unless you are intentionally adjusting the compatibility snapshot.
• Prefer updating the structured daily JSON or graph JSON when building new features.
• If both Markdown and JSON exist for the same day, trust the JSON as the machine-readable source.

Auto-consolidation

Archived conversations are automatically summarized into daily memory, graph snapshots, and weekly plans when enough recent data exists. You usually do not need to manage this manually.