Self-learning context routing for AI coding assistants.
Hologram Cognitive Pro is an MCP server that sits between your AI assistant and your codebase. Works with any MCP-compatible client — Claude Code, Cursor, Cline, Windsurf, Zed, and more. It learns which files you need before you ask — and gets dramatically smarter with every interaction.
Join the WaitlistHologram replaces guesswork with learned, physics-based context routing:
Files carry continuous “attention pressure” that activates, propagates along the code graph, and decays naturally — exactly like real human focus.
Stay deep in one area and it becomes “sticky.” Files you’ve worked in for 5+ turns decay 2.5× slower.
Real AST parsing builds a live symbol graph of function calls, imports, inheritance, and documentation bridges — so vague queries instantly resolve to exact source files.
A tiny 1.08M-parameter set-attention transformer with online learning (F1: 0.972). Pure NumPy, runs on CPU, learns your project’s unique patterns in just ~32 KB of state.
Per-file learned decay/boost parameters that evolve with your actual usage.
While idle, it quietly walks the code graph, pre-warms clusters, and discovers connections you haven’t made yet.
Pulls relevant past decisions and discoveries from your session history with 100% Recall@5.
Multi-machine coordination. Gaps? It auto-dispatches research to the best node (right GPU, right codebase) and returns findings in ~50 seconds — zero human intervention.
When you finish a task, it automatically captures what mattered and links it into long-term memory.
One tool call per turn. That’s it.
hologram_route(query)
Returns ranked files + bridged code + relevant memories
Use the routed context to build, fix, or explore
hologram_observe(files_used)
Learning loop closes — the system gets smarter
| Metric | Value |
|---|---|
| Reranker F1 | 0.972 |
| Reranker inference | ~135 ms CPU / 130 files |
| Memory query | 54 ms |
| Online learning state | 32 KB per project |
| Dependencies | numpy (req), sentence-transformers (opt) |
Limited beta — join the waitlist for access.
We’ll review your application and send a checkout link to your email when a spot opens.
Private, paid access. Shipping today as a full MCP server (11 tools + 2 resources).
Installable package and one-click setup coming soon.
Proudly beta tested by Anima Lux Labs (Sweden) — builders of local-first, privacy-native AI infrastructure.