Save current session's decisions, findings, and context to MemCP persistent memory. Use this skill when the user says "save memory", "save context", "remember this", before compact, at session end, or when you've accumulated significant new knowledge (decisions made, bugs found, preferences learned) that should persist to future sessions. Also trigger proactively when context is getting full or before any destructive context operation.
Persist this session's important knowledge to MemCP so it survives across sessions.
Anything not saved to MemCP is lost when the session ends or context compacts. Decisions get re-debated, bugs get re-discovered, and the user has to re-explain their preferences. Saving proactively prevents this waste.
Scan the current conversation for unsaved knowledge. Look for:
| Type | category | importance | Example |
|---|---|---|---|
| Architecture/design decisions | decision | high or critical | "Use auxiliary loss instead of replacing value target" |
| Bug discoveries and fixes | finding | high | "config missing dynamic attributes causes crash" |
| API gotchas or surprises |
finding |
medium |
"env.legal_actions is a property, not a method" |
| Performance observations | finding | medium | "Solver query < 500ms when remain_count <= 8" |
| User preferences learned | preference | high | "Code comments must be in Traditional Chinese" |
| TODOs identified | todo | medium | "Phase A infrastructure not yet started" |
| Experiment results | finding | high | "B-w0.5 achieved 12% MAE reduction" |
For each item worth saving, call:
memcp_remember(
content="Concise but complete description",
category="decision|finding|preference|todo",
importance="low|medium|high|critical",
tags="comma,separated,keywords",
entities="key_entity1,key_entity2"
)
Tag guidelines:
solver, mcts, config, learner)change-phase, solver-supervision)Avoid saving:
memcp_search(query) first if unsureIf the session produced substantial analysis, experiment results, or summaries:
memcp_load_context(
name="descriptive-name-YYYY-MM-DD",
content="The large content to store"
)
Good candidates: code review results, experiment analysis, architectural comparisons, debug session logs.
Call memcp_status() to verify storage, then tell the user what was saved:
Saved to MemCP:
- 2 decisions (solver auxiliary loss design, curriculum approach)
- 1 finding (config dynamic attribute gotcha)
- 1 context stored:
experiment-analysis-2026-03-08- Total insights: 12 (+3 new)
Save without being asked when:
/compact to the user