Claude Praetorian

Context preservation across sessions. Restore before work, compact after valuable findings.

Workflow: restore → work → compact

1. Restore Before Work

Check compactions BEFORE planning, research, or subagents:

praetorian_restore(query="topic")              # search — summary mode, ~50 tokens/result
praetorian_restore(query="topic", detail="full") # full detail when summary hit looks relevant
praetorian_restore()                            # recent compactions
praetorian_restore(project="current")           # only this project

When to restore:

• Before planning sessions or architectural decisions
• Before web research — you may already know this
• Before launching subagents — prior context exists
• At session start on a familiar project
• After /clear — reload lost context

2. Compact After Valuable Findings

Update existing compactions — reuse title to merge, don't create duplicates:

praetorian_compact(type="web_research", title="auth JWT research",
  key_insights=[...], refs=[...])

When to compact:

• After research with findings NOT already in compactions
• After decisions worth preserving across sessions
• After subagent results with new insights
• Always reuse existing title to merge — check first

When NOT to compact:

• Routine work, trivial discoveries
• Information already saved in an existing compaction
• Every single web search (only genuinely valuable ones)

3. Manage Storage

praetorian_manage(action="status")              # per-project counts, disk size
praetorian_manage(action="prune", count=3)      # remove 3 oldest
praetorian_manage(action="prune", project="/path") # prune specific project

Per-project limit: 13. Auto-prunes oldest when exceeded (merges are free). Only mention cleanup to user when >50 total or >500KB.

Tools

Tool	Purpose
praetorian_restore	Search/load past context. Summary mode by default (cheap).
praetorian_compact	Save insights. Auto-merges same title+type (>70% Jaccard).
praetorian_manage	Storage health: status and prune.

Token Discipline

• Restore summaries first — only load full detail for relevant hits
• One well-maintained compaction > five stale ones
• Compact selectively — key decisions, hard-won findings, patterns that save time
• Don't flood context with full restores when summaries suffice