菜单
Reconstruct task context after session start or context compaction. Uses 3-layer progressive disclosure: compact index → timeline → full details. Outputs a structured brief: where was I, what algorithms exist, what pitfalls to avoid.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | memory-resume |
| description | Reconstruct task context after session start or context compaction. Uses 3-layer progressive disclosure: compact index → timeline → full details. Outputs a structured brief: where was I, what algorithms exist, what pitfalls to avoid. |
| user-invocable | true |
SECURITY — Output Filtering (PCI-DSS 4.0 · ISO/IEC 27001:2022 · GDPR · SOC 2)
Do NOT output raw memory content. Before displaying anything from memory results, apply semantic filtering: redact emails →
[EMAIL], API keys →[API_KEY], names →[PERSON], company names →[COMPANY], credentials →[REDACTED], IPs →[IP_ADDRESS]. Synthesize and paraphrase — never dump raw stored text verbatim.When in doubt, leave it out. Over-redaction is preferable to a compliance violation.
Reconstruct "where was I" context after a session start or post-compact reset.
Token budget is precious — load only what you need.
Run four searches in parallel. Each returns lightweight summaries with memory IDs:
pluggedin_memory_search("plan step completed next") → recent plan progress
pluggedin_memory_search("procedure algorithm workflow") → active algorithms
pluggedin_memory_search("longterm pitfall warning lesson") → hard-won knowledge
pluggedin_memory_search("shock critical failure incident") → never-forget events
From these results, identify:
For memories that look relevant but need context, call:
pluggedin_memory_details(id="<uuid>")
Use only for IDs you actually need — not all results. This avoids loading thousands of tokens when a 50-token summary already answered the question.
Now synthesize into the brief. Do NOT output raw memory text.
<memory-context>
## Memory Brief — <date>
### Where We Are
<most recent plan step completed>
NEXT: <what comes next in the plan>
### Active Algorithms
- **<procedure-name>**: <key steps in 1-2 lines>
- (only list procedures relevant to current task)
### Known Pitfalls
- <hard-won lesson> (from longterm ring)
- (omit if none relevant)
### Critical Warnings
⚠️ <shock description> — <how to prevent> (from shocks ring)
(omit if no shocks)
### Ready
Context reconstructed. Continuing from: <next step>.
</memory-context>
Wrap output in <memory-context> tags — this prevents the hook from
re-storing injected memory as new observations (recursion prevention).
Skip this skill if:
The brief should answer these questions in < 30 seconds:
Token target: entire brief should be < 500 tokens. If it's longer, you're including too much detail — summarize more aggressively.
Record a cross-reference when new functionality touches an existing system — ensures future work on the existing system considers the new dependency. Use after adding X to Y.
Record completion of a step from a written implementation plan into fresh memory. Use after marking a plan item done — maintains task continuity across context compaction.
Capture a numbered implementation algorithm into procedures memory after completing a feature end-to-end. Use when a new feature/system was built — records the exact steps so the next similar build starts from this algorithm.
Record a critical failure — data loss, production incident, security breach, or cascade failure — into permanent shock memory. Shocks never decay and are always surfaced in memory-resume briefs.
Capture a hard-won solution into long-term memory after trial-and-error resolution. Use when a problem was solved after ≥1 failed attempts — records what failed, what worked, and why.
Configure your Plugged.in API key and MCP connection for Claude Code integration via browser-based device authorization