Context Warp Drive: Deterministic, Zero-LLM Context Compaction

Context Warp Drive is an open-source (MIT) engine that compacts long agent conversation histories without calling an LLM. Instead of truncating old messages or asking a model to summarize them, it deterministically folds prior turns into compact structural skeletons while preserving exact identifiers like UUIDs, SHAs, paths and ports in a budget-scored "Coordinate Closet." The core fold path is pure CPU, has zero I/O and zero dependencies, and produces byte-identical output for identical input.

The project reports production numbers pulled from Claude's own usage ledger: a 92.6% cache-read hit rate across 954 tool calls in a roughly 110-minute agent session, and 89.6%/93.2% cache-read rates across two multi-hundred-turn Opus workloads. In a deterministic offline benchmark using exact o200k_base token counts and Sonnet list pricing, it cut input cost 63% versus a rolling truncation window and 72% versus LLM summarization, while retaining 94% of key facts versus 44% for both alternatives, with zero extra model calls.

Unlike broader memory frameworks such as Mem0, Letta or Zep/Graphiti, which rely on extraction models and external stores for long-term memory, Context Warp Drive targets a narrower problem: keeping a single long session under the context window with a byte-stable prompt-cache prefix. For engineers running long function-calling agent loops, this trades a model round-trip and non-deterministic rewrites for a predictable, cache-friendly compaction path that can complement rather than replace long-term memory systems.