AJNT Token Saving Proxy

ActiveContext is a solution designed for long-running coding agents that prevents performance degradation caused by context bloat. By sitting between the agent and the LLM, it actively manages session history, prunes irrelevant data, and preserves critical task information to ensure agents remain focused and efficient.

Key Features

• Between-task pruning: Automatically removes stale tool calls, abandoned plans, and redundant file reads.
• Goal-aware preservation: Identifies the primary objective and keeps essential reasoning and state information intact.
• Cache-aware architecture: Designed to work with providers offering explicit prompt caching (e.g., Anthropic, Alibaba Cloud) to maintain high hit ratios.
• Dual-core intelligence: Uses an asynchronous curator layer that runs between turns to restructure history without blocking the primary agent.
• Per-agent tuning: Allows users to adjust the aggressiveness of context management based on specific agent types, models, or workflows.

How it works

ActiveContext uses both input compression (like Headroom) and a secondary intelligence layer to monitor the session as it unfolds. Instead of simple compression, it identifies what the agent is trying to accomplish and curates the context accordingly. When a task completes, it restores exactly the information needed for the next step, resulting in 20–70% fewer tokens used and the avoidance of disruptive compaction events.