Runtime Modes

TokenPilot provides three runtime modes: conservative, normal, and aggressive. Each is a preset that configures the stabilizer, reduction, and eviction subsystems. Switch modes at any time without restarting.

Mode Comparison

Behavior	Conservative	Normal	Aggressive
Stable Prefix	On	On	On
Stabilizer target	Developer	Developer	User
Reduction	Light	Balanced	Strong
Reduction trigger	4000 chars	2200 chars	1400 chars
Max tool chars	1800	1200	900
Eviction	Off	Off	On
Risk of signal loss	Very low	Low	Moderate
Token savings	Moderate	High	Maximum

Conservative Mode

For safety-critical sessions where you cannot afford to lose any context.

lightmem2 mode conservative

• Stabilizer rewrites context into cache-stable form but keeps it developer-visible
• Reduction applies light trimming only — removes truly redundant output
• Eviction is disabled — full history is always available

Normal Mode

The default and recommended mode for most sessions.

lightmem2 mode normal

• Stabilizer rewrites context and attaches dynamic content at the developer level
• Reduction applies balanced trimming — removes noise while keeping signal
• Eviction is disabled by default — for most sessions, reduction alone is sufficient

Aggressive Mode

For maximum savings when you're willing to trade some context completeness.

lightmem2 mode aggressive

• Stabilizer rewrites context and attaches dynamic content at the user level (further from the model)
• Reduction applies strong trimming — may discard borderline-useful output
• Eviction is enabled with lower thresholds — older context is dropped sooner

Switching Modes Mid-Session

You can change modes at any time — no restart needed:

# Per-host (OpenClaw supports all three modes)
lightmem2 openclaw mode aggressive

# Codex and Claude Code only support conservative and normal
lightmem2 codex mode conservative
lightmem2 codex mode normal
lightmem2 claude-code mode conservative
lightmem2 claude-code mode normal

# Or inside OpenClaw session
/lightmem2 mode aggressive

The new mode takes effect on the next turn.

Next

• Stable Prefix — how stabilization works
• Context Reduction — the trimming pipeline
• Context Eviction — lifecycle-aware pruning