crux-recall

Prompt file imported from zotoio/CRUX-Compress (.cursor/commands/crux-recall.md). Fill in {{arguments}} before use. Copyright stays with the author.

crux-recall

Decompress, query, and display CRUX memories in human-readable form.

Repository: github.com/zotoio/CRUX-Compress

Usage

/crux-recall                                    - Show contextually relevant memories
/crux-recall "search query"                     - Search memories by keyword
/crux-recall 20260403-crux-memories              - Show memories from a specific spec
/crux-recall memories/core/some-memory.memory.md - Display a specific memory file
/crux-recall --total                             - Visualize the entire memory system as an interactive canvas

Parameters

Instructions

When this command is invoked, spawn a crux-cursor-memory-manager subagent in Recall mode to query and display memories. Recall is read-only — it never modifies memory files on disk.

User input escalation — CRITICAL: This command uses Pattern B (work first, then escalate) — the subagent queries and formats memories first, then the parent handles post-display interactions. The subagent NEVER calls AskQuestion directly. After displaying recall results, the parent agent (you) handles all post-display interactions (delete, consolidate, promote, skip) using AskQuestion. If the subagent needs user input, it returns a needs_user_input section in its response.

Argument Handling

• No arguments: The manager loads the memory index (.crux/memory-index.yml) and surfaces memories most likely to be relevant to the current conversation context. For each memory, it shows title, type, strength, reference count, and a brief rationale for why it was surfaced.
• Quoted text (e.g. "performance optimization"): The manager searches existing memories by title, description, tags, and body content. Results are ranked by relevance with decompressed body content shown for compressed memories. Pass {{arguments}} to the subagent as the search query.
• Spec name(s) (e.g. 20260403-crux-memories): The manager finds all memories whose source field matches the given spec slug(s). Results are grouped by type. Pass {{arguments}} to the subagent as the spec name(s).
• File path(s) (e.g. memories/learning/foo.memory.md): The manager reads the specified memory file(s) directly. Compressed files (.memory.crux.md) are decompressed for display. Full frontmatter and body are shown. Pass {{arguments}} to the subagent as the file path(s).
• --total: The manager gathers the complete memory corpus — index metadata, all memory files, and their relationships — then uses /canvas to generate an interactive visualization of the entire memory system using only cursor/canvas SDK components and inline SVG. No table or text output is produced; the canvas is the deliverable.

What Happens

1. The manager reads .crux/crux-memories.json to load configuration
2. Based on the invocation mode, it either:
   • Loads the memory index and selects relevant entries
   • Searches memory files for keyword matches
   • Filters memories by source spec slug
   • Reads specific memory files directly
   • (--total) Reads the full memory index and all memory files, builds a relationship graph, and generates a canvas visualization using cursor/canvas SDK primitives and inline SVG
3. For compressed memories (.memory.crux.md), decompresses the CRUX body to terse natural language for display — without modifying the file on disk
4. Presents results with frontmatter metadata and readable body content (or as an interactive canvas when --total is used)

Display Format

Render memories as a markdown table grouped by type. Each row shows the memory's short hash id for easy reference.

Table format (one table per type group):

### {Type} Memories

| ID | Title | Str | Refs | Source | Tags |
|----|-------|-----|------|--------|------|
| `{id}` | {title} | {strength} | {references} | {source} | {tags as comma-separated} |

After the table(s), show a Details section with each memory's body content (decompressed if needed), prefixed by its hash:

#### `{id}` — {title}

{body content — decompressed if needed}

Post-Display: Next Steps Menu

After displaying the Recall results, use the AskQuestion tool with a single multi-select question offering these actions:

If the user selects one or more actions, execute them in order:

1. Delete: Present a follow-up AskQuestion listing each memory by its id and title as a multi-select. Delete selected memories using crux-skill-memory-crud Delete, then rebuild the index via crux-skill-memory-index.
2. Consolidate: Spawn a crux-cursor-memory-manager subagent in REM Sleep mode scoped to the displayed memories. The subagent identifies near-duplicates, merges them, and compresses if enableMemoryCompression is active. Present the consolidation report to the user.
3. Promote: Present a follow-up AskQuestion listing each memory by its id and title as a multi-select. For selected memories, increment strength by 1 via crux-skill-memory-crud Update, then rebuild the index.
4. Skip: End the interaction with no further action.

Related

• crux-cursor-memory-manager agent — The specialist that manages the memory lifecycle
• crux-skill-memory-compress skill — Decompression logic for compressed memories
• crux-skill-memory-index skill — Memory index used for discovery
• /crux-dream — Extract and create memories from completed work
• /crux-forget — Remove memories from the corpus
• /crux-remember — Create ad-hoc memories outside of spec workflows
• /crux-meditate — Recursive memory-informed exploration