Memory: Why Long Conversations Don't Break It

Background

Claude has a token limit (~200k). On complex tasks that run long, conversation history can overflow the context window. Three strategies address this problem.

Threshold System

Context window (e.g. 200k)
  └── Minus output reservation (20k) = effective window (180k)
        ├── Minus buffer (13k) = auto-compact trigger threshold (167k)
        ├── Minus (20k) = warning threshold (shows token remaining warning)
        └── Near limit = blocking threshold (stops accepting new input)

Three Strategies

1. autoCompact — Automatic Compression (primary strategy)

• Trigger: Token usage reaches the auto-compact threshold
• What it does: Uses a separate Claude API call to summarize conversation history into a digest, replacing the old records
• Abuse prevention: Stops after 3 consecutive failures (circuit breaker)
  • Reason: Stats show 1,279 sessions failed consecutively 50+ times, wasting ~250,000 API calls per day
• User-controlled: Can be disabled via settings or DISABLE_AUTO_COMPACT env variable

2. reactiveCompact — Reactive Compression (error recovery)

• Trigger: Specific API errors (max_output_tokens truncation, prompt_too_long overflow)
• What it does: On error, compresses context silently before retrying — invisible to the user
• Status: Feature Flag REACTIVE_COMPACT, supplements autoCompact

3. contextCollapse — Context Collapse (experimental, aggressive)

• Trigger: Feature Flag CONTEXT_COLLAPSE enabled
• What it does: More aggressive than compression — lets Claude actively decide via CtxInspectTool which content can be folded and discarded
• Design philosophy: Lets Claude actively manage its own memory, similar to humans "selectively forgetting unimportant details"

How They Relate

During normal operation
  → tokens gradually grow
    → approaching threshold → [autoCompact] summarizes history, continues running

On API error
  → max_tokens / prompt_too_long → [reactiveCompact] silently compresses then retries

Context too complex (experimental)
  → [contextCollapse] Claude autonomously decides what to fold away