Summary

Add built-in context management and memory optimization features to reduce token costs and prevent context bloat.

Problem

Long-running agents accumulate context through:

• Growing memory files (daily logs, learnings, decisions)
• Skill descriptions loaded every turn
• Cron job payloads and histories
• Session transcripts

This leads to:

• Increased token costs ($5-10/day for active agents)
• Context window pressure requiring frequent compaction
• Manual maintenance burden on users

Proposed Features

1. Automatic Memory Archiving

• Auto-archive memory files older than N days (configurable)
• Compress archived files (gzip)
• Keep archived content searchable via memory_search
• Config: agents.defaults.memory.archiveAfterDays: 7

2. Context Budget Monitoring

• Track context file sizes
• Alert when approaching threshold
• Dashboard widget showing context usage breakdown
• Config: agents.defaults.context.maxKb: 50

3. Smart Memory Summarization

• Auto-summarize old daily logs before archiving
• Extract key learnings/decisions to dedicated files
• Preserve important context while reducing size

4. Lazy Skill Loading

• Load skill descriptions on-demand instead of every turn
• Cache recently used skills
• Reduce base context by 3-5KB

5. Memory Maintenance Hooks

• onMemoryMaintenance hook for custom logic
• Built-in weekly maintenance cron
• Dry-run mode for testing

Example Config

{
  "agents": {
    "defaults": {
      "memory": {
        "archiveAfterDays": 7,
        "compressArchive": true,
        "autoSummarize": true,
        "maxActiveKb": 50
      },
      "context": {
        "lazySkillLoad": true,
        "budgetAlertKb": 40
      }
    }
  }
}

Benefits

• Reduces daily token costs 20-40%
• Prevents runaway context growth
• Zero manual maintenance
• Preserves all learnings (archived, not deleted)

Workaround

Currently using a bash script for weekly maintenance - happy to share if helpful for implementation reference.

Submitted by @sachah via Alex agent