Research


Similarity search returns what is close, not what is related. HydraDB connects your context into a structured, temporally versioned graph and gives agents the exact context they need: relational first, preference aware, and precise. This is where we share how we build that context layer, and how it holds up on public, reproducible benchmarks.

The thesis

Similarity is not relevance.

A longer context window does not make an agent stateful. Real continuity comes from structure. HydraDB stores knowledge as a versioned graph that preserves every state transition, enriches each fact so it stands on its own, and fuses semantic, lexical, and relational signals at recall time.

When context is structured at the database level, answer quality stops depending on raw model size. The same architecture reaches state of the art results from a compact GPT-5 Mini to Gemini 3.0 Pro.

All research

Topics of our research

Versioned knowledge graph
Knowledge as an append-only ledger. Updates commit new timestamped edges instead of overwriting, so no history is lost and every past state stays addressable.
Temporal reasoning
Valid-time metadata on every edge lets agents resolve when a fact became true, when it changed, and what holds now, without confusing stale records for current ones.
Sliding-window inference
Lightweight enrichment resolves pronouns and references against a look-back and look-ahead window, turning ambiguous fragments into self-contained, retrievable facts.
Hybrid and graph retrieval
A multi-stage pipeline fuses dense, sparse, and graph signals through query expansion, entity traversal, and cross-encoder reranking to recover the correct factual and relational state.
Tiered storage and decay
A retention score blending salience, recency, and reuse moves context across hot memory, SSD, and object storage, so high-signal records stay fast and noise is archived cheaply.
Model-agnostic context
We study how much of an agent's competence can move out of the model and into the context layer, and we measure stability across backbone models of very different scale.

Build AI with compounding intelligence.