Summary

Add context window budget management to execution loops — soft budget indicators, compaction hooks at task boundaries, and compressed checkpoint recovery.

Motivation

Three independent research sources converge on the same need: agents must manage their context window actively during long-running tasks.

1. Memex (arXiv:2603.04257): Soft budget indicators — append token usage status so agent self-manages compression timing
2. LangChain Autonomous Context Compression: Agent-triggered compaction at clean task boundaries (conservative tuning)
3. KV Cache Attention Matching (arXiv:2602.16284): Compress-on-fill pattern — 6 consecutive online compressions with no quality loss

Design

Soft Context Budget Indicators

Append [Context: 12,450/16,000 tokens | 3 archived blocks available] to agent context. Works with existing TurnRecord token counting. Agent self-manages compression timing at natural task boundaries.

Compaction Hook

Add a hook point in execution loops where the agent (or the loop itself) can trigger context summarization between turns. Detect context window fill level and optionally summarize older turns.

Compressed Checkpoint Recovery

When recovering from a checkpoint, inject compressed summary + index table from last compression point instead of replaying full conversation history (from Memex segmented trajectory pattern).

Affected Files

• src/ai_company/engine/react_loop.py
• src/ai_company/engine/plan_execute_loop.py
• src/ai_company/engine/checkpoint/ (recovery with compressed context)
• src/ai_company/engine/prompt.py (inject budget indicators)

Research

• Memex Indexed Experience (arXiv:2603.04257)
• LangChain Autonomous Context Compression
• KV Cache Attention Matching (arXiv:2602.16284)