eatakishiyev/context-forge

ContextForge is a tool that sits between your AI agent and the language model it calls, cleaning up the conversation history before each request is sent. The problem it addresses is called context rot: as an agent's conversation grows longer, the model's accuracy can drop 30 to 50 percent well before the model's official context window limit is reached. This happens because important facts get buried in the middle of a long history, duplicated content adds noise, and the model's attention weakens on older material. The tool works as a four-step pipeline. It scores each request with a 0 to 100 rot risk number broken into four parts: load, redundancy, middle burial, and fragmentation. It compresses the history by removing near-duplicates and stale low-importance content using extractive methods that never paraphrase away a key fact. It reorders content so the most important information appears at the start and end of the context, where attention is strongest. Finally, it enforces a hard token budget by dropping the least important content first, with every drop recorded in an audit log. You can use it as a Python library (pass a Trace object through a ContextCompiler) or as a drop-in proxy: point your Anthropic or OpenAI SDK's base URL at the local proxy server and it compiles every request transparently without changing your existing code. A benchmark harness lets you measure accuracy improvement and token savings on your own traces. The tool also includes a security scanning module that inspects prompts and tool results for leaked API keys, PII, and prompt injection, and can redact or block traffic at the gateway level. This is for developers building AI agents that run long multi-step sessions and are seeing unexplained accuracy drops or higher-than-expected token costs.