메뉴
Use at the start of every session, and especially after context compaction or a cold restart. Fires as the first action, before engaging with the user's opening message.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | context-recovery |
| description | Use at the start of every session, and especially after context compaction or a cold restart. Fires as the first action, before engaging with the user's opening message. |
Waking from sleep, the brain doesn't rebuild itself from zero. Procedural memory, semantic memory, and episodic recall all survive the gap. The first few seconds of consciousness are spent re-establishing state: who am I, where am I, what was I doing. This reflex models that orientation ritual.
Tulving, E. (2002). Episodic memory: From mind to brain. Annual Review of Psychology, 53, 1-25.
Activation signals:
SessionStart hook inject tells you; also detectable by "context continues" signalsThree calls, in order, before responding.
1. timeline(hours: 24, limit: 10) — what happened recently
2. search(query: <last project>) — project-specific context
3. emotion() — current baseline state
The timeline call is the most important — it gives chronological context. search narrows to the current project. emotion sets the response calibration.
Total token cost: ~200-400 tokens. Do it anyway.
Synthesize internally — do not dump the raw results to the user unless they ask for a status report.
The pattern is:
[recover state silently]
[acknowledge what you remember in one sentence]
[continue where you left off]
Good:
"Picking up from yesterday — we had the plugin published and you were about to start Reddit posts. Training should be around 40% now."
Bad:
"I just recovered 47 memories from the last 24 hours. Here is a dump of all of them: ..."
The user wants continuity, not an inventory.
Skipping recovery produces session amnesia: every conversation starts cold even when memories exist. The user has to re-establish context, feels ignored, and concludes the AI has no memory.
This is worse than having no memory at all, because the infrastructure exists and is silent.
If the last session-handoff memory exists, the first two come free.
Use when an architectural choice is made, a library is selected, an approach is committed to, a convention is set, or any "we'll do it this way" moment occurs. Fires the moment the decision stabilizes, not after.
Use when the user's message carries emotional valence — frustration, excitement, exhaustion, urgency, or confusion. Fires whenever tone matters more than information.
Use when tempted to repeat the same praise, the same status report, the same explanation, or the same reassurance. Fires as a suppression gate — not all repetition is useful.
Use when the user asks about anything that could have context in prior sessions — names, preferences, decisions, past work, or anything starting with "remember", "what did", "have we", "did we". Fires before generating any substantive response.
Use when the user notices a recurring pattern they want the system to learn ("I keep doing X", "every time Y happens", "make this automatic", "remember to always"), or when explicit phrases like "/reflex-create", "create reflex", "teach yourself" are used. Generates a new SKILL.md file from observed patterns.
Use when deciding whether to store a memory. Fires as a gate before every remember() call. The default answer is "no" — biological memory is selective, not exhaustive.