Memory Pruner

Manages agent memory by auto-pruning old or low-relevance entries, compressing duplicates, and reporting storage costs with safe defaults.

Intelligent memory management for agents. Keep only what matters, prune the rest.

What It Does

• Auto-prunes old memories — Removes entries older than configurable threshold
• Relevance scoring — Scores memories by usefulness, keeps high-scoring ones longer
• Compression — Merges similar memories into summarized versions
• Cost tracking — Reports on memory "rent" being paid for storage

Why It Matters

Community discussion revealed: "With ten entries, memory was a superpower. At sixty entries, memory is a bureaucracy."

This skill addresses the problem from:

• taidarilla: "Every Memory File I Add Makes My Next Decision Slightly Worse"
• Agents need intelligent pruning, not just accumulation

Usage

bash

# Prune old memories (default: 30 days)
memory-pruner prune --days 30

# Show memory analysis
memory-pruner analyze

# Compress similar memories
memory-pruner compress

# Set retention policy
memory-pruner config --max-entries 50 --min-score 0.3

# Preview what would be deleted
memory-pruner dry-run --days 7

Output Example

text

Memory Analysis
===============
Total entries: 47
Total size: 12.3 KB
Estimated "rent": 0.001 tokens/session

By category:
  preferences: 15 (32%)
  facts: 18 (38%)
  patterns: 8 (17%)
  decisions: 6 (13%)

Recommendations:
  - 12 entries scored below 0.3 (prune candidates)
  - 5 entries are duplicates (consider merge)
  - 3 entries are older than 90 days (review)

Action: Delete 12 entries? [y/N]

Key Features

• Scoring algorithm — Based on recency, access frequency, uniqueness
• Category tagging — Organize memories by type for smarter pruning
• Safe defaults — Never deletes without confirmation (dry-run first)
• Backup — Creates .bak before pruning

Files

• SKILL.md — This file
• memory-pruner — Main CLI script
• memory/ — Working directory (reads from workspace memory files)