HoloCore trims AI input tokens with a local, layered context system
HoloCore is a system that stops AI tools from sending the entire project, memory store, or on every request. Instead it organizes project knowledge into three layers. Atlas maps project structure, components, and relationships. Archive stores curated, durable knowledge.
Animus holds episodic history and past decisions. For each new request, it first picks the relevant route: a code or structure question starts with Atlas, Archive is added only when documented knowledge applies, and Animus is added only when prior decisions or matter. Only the selected context is passed to the connected AI client through CLI or MCP . The system is designed to be — the underlying AI model can be swapped while the project map, curated knowledge, , and locally stored data stay in one installation.
It also avoids feeding its own output back into itself, preventing s. In a local benchmark using five project questions, HoloCore used roughly 156 estimated tokens.
Key points
- Splits context into three layers: Atlas (structure map), Archive (curated knowledge), Animus (history/decisions)
- Selectively includes only the layers relevant to each request
- Connects to AI clients via CLI/MCP without being tied to one model
- Avoids feeding its own output back in, preventing s
- Used about 156 estimated tokens on a 5-question local benchmark (comparison baseline unclear)