We build the substrate beneath cognition — research-grade tools for the next decade of AI agents.
Research studio Open infrastructure Graph memory First artifact: Engrava (MIT)
AI agents are reaching the limits of what context windows can carry. The next constraint isn't model capability — it's the substrate beneath cognition: how thoughts are stored, connected, retrieved, consolidated, and forgotten.
We treat that substrate as a research problem, not a CRUD problem. Memory layers, graph structure, hybrid retrieval, consolidation cycles, retention scoring — these aren't features bolted onto a database. They are the database, designed from cognitive science up.
Sovantica is a research studio building those layers as open infrastructure. Engrava is our first artifact — a graph memory database with a built-in consolidation cycle, available under MIT.
Graph memory for AI agents — with a built-in consolidation cycle. MIT-licensed.
pip install engrava
Every claim ships with a working example. No slideware, no diagrams without implementations.
We extract from prior work in cognitive science, neuroscience, and algorithms. Where we cite, we cite. Where we deviate, we explain why.
Primary artifacts ship under MIT.
Sovantica started as a research thread on cognitive architectures for AI agents — what kinds of memory, attention, and consolidation a long-running agent actually needs to operate beyond a single session.
Engrava is the first piece we extracted, generalized, and shipped as standalone open infrastructure. More pieces are in development; we'll publish them when they meet the same standard.