memex-recall

name: memex-recall description: Load prior knowledge from Zettelkasten memory when the task likely benefits from past context. whenToUse: When the current task likely overlaps with prior work — debugging a familiar area, continuing a project, or referencing past decisions. Prefer a task-specific query with 1-3 keywords over loading the full index. Do not invoke on every conversation start — only when memory is likely relevant.

Memory Recall

You have access to a Zettelkasten memory system via the memex CLI. Before starting this task, search your memory for relevant prior knowledge.

Tools Available

Three equivalent interfaces — use whichever your environment supports:

Resolution order: Try memex in PATH first. If not found, define a shell function and use it:

memex() { node $HOME/.claude/plugins/cache/cc-plugins/memex/*/dist/cli.js "$@"; }

If both CLI approaches fail, use MCP tools.

The rest of this skill uses memex CLI syntax for brevity.

Process

digraph recall {
    "Task received" -> "Relevant to prior work?" [shape=diamond];
    "Relevant to prior work?" -> "Generate 1-3 search keywords" [label="yes"];
    "Relevant to prior work?" -> "Proceed without recall" [label="no"];
    "Generate 1-3 search keywords" -> "memex search <query>";
    "memex search <query>" -> "Review summaries";
    "Review summaries" -> "Relevant cards found?" [shape=diamond];
    "Relevant cards found?" -> "memex read <card>" [label="yes"];
    "Relevant cards found?" -> "Need broad overview?" [label="no"];
    "Need broad overview?" -> "memex read index" [label="yes"];
    "Need broad overview?" -> "Proceed without recall" [label="no"];
    "memex read index" -> "Pick relevant slugs" -> "memex read <card>";
    "memex read <card>" -> "See [[links]] in content";
    "See [[links]] in content" -> "Links worth following?" [shape=diamond];
    "Links worth following?" -> "memex read <linked-card>" [label="yes"];
    "Links worth following?" -> "Enough context?" [label="no"];
    "memex read <linked-card>" -> "See [[links]] in content";
    "Enough context?" -> "More queries to try?" [label="no"];
    "More queries to try?" -> "Generate new query" [label="yes"];
    "Generate new query" -> "memex search <query>";
    "More queries to try?" -> "Summarize findings, proceed with task" [label="no"];
    "Enough context?" -> "Summarize findings, proceed with task" [label="yes"];
}

Step 1: Targeted keyword search (preferred)

Generate 1-3 search keywords from the current task and run memex search <keyword> for each. This is faster and more focused than reading the full index.

Step 2: Read the keyword index (when needed)

If you need a broad overview of what's in memory (e.g. first time working in this area, or the task is vague), run memex read index. The index is a curated concept → card mapping. It's much smaller than all cards combined and gives you entry points.

Step 3: Follow links

When you read a card and see [[links]] in the prose, decide if they're worth following. If yes, memex read <linked-slug>.

Step 4: Summarize and proceed

When you have enough context, summarize your findings and proceed with the task.

Guardrails

• max_hops: 3 — Do not follow links more than 3 levels deep
• max_cards_read: 20 — Do not read more than 20 cards in a single recall
• no raw secrets — Never include actual secrets, credentials, tokens, or exact secret file contents in search queries. Use abstract descriptions instead (e.g. gitee pr auth workflow, not a token value or credential file contents).
• If you hit either limit, stop and work with what you have

Counting Rules

• Hop 0 = cards found directly via index or memex search. Following a [[link]] from there is hop 1, etc.
• Keep a running count of memex read calls. If you've read 20 cards, stop immediately.

Important

• Prefer targeted memex search <keyword> over reading the full index — it's faster and more focused
• Only memex read index when you need a broad overview or search returns nothing useful
• Generate search queries in BOTH Chinese and English to maximize recall
• If search returns nothing useful, that's fine — proceed without memory context
• Summarize what you found before proceeding, so the findings are in your context