AgentSkillsCN

continuation-recovery

在会话延续时，通过有针对性的 episodic memory 搜索恢复上下文。当 Claude 检测到会话延续摘要、引用了先前未纳入当前上下文的工作内容，或用户说出“从上次停止的地方继续”时，该功能会自动触发。涵盖会话延续、上下文恢复、情景记忆。

SKILL.md
--- frontmatter
name: continuation-recovery
description: Recover context at session continuations using targeted episodic memory searches. Auto-triggered when Claude detects a continuation summary, references to prior work not in context, or the user saying "continue where we left off". Covers session continuation, context recovery, episodic memory.
user-invocable: false

Continuation Recovery

Recover key decisions, gotchas, and preferences from prior sessions WITHOUT flooding the context window. This skill is auto-triggered -- the user does not invoke it.

When to Trigger

Activate when ANY of these signals are present:

  • •A conversation summary block mentioning "continued from a previous conversation"
  • •User says "continue", "pick up where we left off", or similar
  • •References to completed work that is not visible in the current context
  • •The conversation starts with a task context but no visible history

Instructions

1. Gather search terms (do not search yet)

From the continuation summary and environment, collect:

  • •Project name: from the working directory path (e.g., "chronolog")
  • •Branch name: from git branch --show-current
  • •Task keywords: 2-3 specific nouns from the summary (e.g., "offline sync", "conflict resolution")

2. Search episodic memory

Use the mcp__plugin_episodic-memory_episodic-memory__search tool with:

  • •query: an array of 2-4 specific terms (project name + branch or task keywords)
  • •limit: 15
  • •mode: "both"

Example:

code
query: ["chronolog", "offline-sync", "conflict resolution"]
limit: 15

If the first search returns fewer than 3 results, run a second broader search using just the project name and a single general keyword (e.g., ["chronolog", "migration"]). Cap at 2 searches total.

3. Extract only actionable items

From the search results, extract ONLY:

  • •Decisions: what was chosen and why (1 line each)
  • •Gotchas: errors hit and their fixes (1 line each)
  • •Preferences: user conventions or stated preferences (1 line each)

IGNORE: code snippets, exploratory discussion, verbose explanations, anything already in CLAUDE.md or MEMORY.md.

4. Update working memory

If findings contain information NOT already in the project's MEMORY.md, append a brief section:

code
## Session Recovery [date]
- Decision: [what] because [why]
- Gotcha: [problem] -- fix: [solution]

Keep additions under 15 lines. If everything is already captured, do not write anything.

5. Resume work

Summarize recovered context to the user in 2-3 sentences, then proceed with their request. Do not recite the full memory contents.