pmm-dump

name

pmm-dump

description

Dump PMM memory state as ASCII art in the terminal. Three depth levels: status (heatmap only), summary (heatmap + clusters + timeline), detailed (full ASCII). Runs as a subagent. Use when the user runs /pmm-dump or asks for a text-based memory overview. Trigger on: "pmm-dump", "/pmm-dump", "dump memory", "ascii memory", "text memory overview", "show memory heatmap", "memory dump", or any request for a text-based ASCII visualization of memory state.

PMM Dump

Render PMM memory state as inline ASCII visualizations. Runs as a subagent to keep the main context clean.

Depth level: $ARGUMENTS (default: status if empty or not provided)

Invocation

• /pmm-dump or /pmm-dump status — heatmap only (status level)
• /pmm-dump summary — heatmap + cluster list + last 5 timeline entries
• /pmm-dump detailed — full ASCII: graph map + heatmap + similarity matrix + clusters

Behaviour

Dispatch a general-purpose agent using the Readonly Agent Model from memory/config.md (default: haiku) with the prompt below. Replace <level> with the depth level (status, summary, or detailed). Replace <project-root> with the actual project root path.

Output the agent's returned string verbatim — it contains the fully formatted ASCII visualization.

Agent Prompt

Render PMM memory state as ASCII visualizations. This is a READ-ONLY task — do not edit any files. You may run git commands for timestamps.

Project root: <project-root> Depth level: <level>

Depth Levels

• status — Heatmap only
• summary — Heatmap + cluster list + last 5 timeline entries
• detailed — Full ASCII: graph map + heatmap + similarity matrix + clusters

Visualization 1: Heatmap — File Activity (all levels)

1. Read <project-root>/memory/config.md to get the list of active files
2. For each active file, run: git log -1 --format="%ar|%at" -- memory/<filename>
3. Map the unix timestamp to a heat level:
   • ████ = modified < 5 minutes ago
   • ███░ = modified < 30 minutes ago
   • ██░░ = modified < 2 hours ago
   • █░░░ = modified < 24 hours ago
   • ░░░░ = modified > 24 hours ago or never
4. Check if file is empty/template-only (only has comments and headers, no real content)
5. Sort by recency (most recent first)
6. Render as aligned table with legend

Token Burn Estimate (all levels)

After the heatmap, show a token burn estimate:

1. Count total characters across all .md files in memory/: cat memory/*.md | wc -c
2. Estimate read tokens: total_chars / 4
3. Get last memory diff stats: git diff HEAD~1 --stat -- memory/ | tail -1 (extract insertions + deletions)
4. Estimate write tokens: (insertions + deletions) * 20 / 4
5. Render:

Token Burn (per save)
  Read:  ~12,400 tok    Write: ~850 tok    Total: ~13,250 tok

Use compact single-line format for dump. Round to nearest 50. Comma-separate thousands.

Visualization 2: Clusters + Timeline (summary and detailed only)

If level is summary:

1. Read <project-root>/memory/vectors.md
2. Parse cluster lines: Cluster: name → [members] | theme: ...
3. Render cluster names with member count
4. Read <project-root>/memory/timeline.md
5. Show last 5 timeline entries

Visualization 3: Graph Map (detailed only)

If level is detailed:

1. Read <project-root>/memory/graph.md
2. Parse all [[A]] → relationship → [[B]] lines
3. Group edges by ## section headers
4. Render each group as an ASCII subgraph:
   • Nodes boxed: ┌──────┐ │ name │ └──────┘
   • Edges: ───relationship──▶
   • Group under section name
   • If >15 edges, show top-level structure only (collapse leaf nodes)
   • If empty/template: "No relationships recorded yet."

Visualization 4: Similarity Matrix (detailed only)

If level is detailed:

1. Read <project-root>/memory/vectors.md
2. Parse similarity lines: [[A]] ↔ [[B]] | score: X.XX | basis: ...
3. Parse cluster lines: Cluster: name → [members] | theme: ...
4. Build sparse similarity matrix
5. Render as aligned ASCII table (abbreviate names to 6 chars, · for no data)
6. Render clusters as tree with ┌├└ chars

Output Format

Return the fully formatted ASCII output. Keep it compact. No preamble — start directly with the visualization.

Notes

• The agent does all file reads and git commands — main context stays clean
• Output the agent's return value verbatim (it's the formatted ASCII visualization)