Deyta
Deyta builds the data layer for long-horizon AI agents through its open-source Python library, Khora. Khora ingests unstructured and structured multi-source data into graph, vector, and relational storage, exposing a single query substrate with temporal-aware memory. Its VectorCypher engine routes each query to the appropriate retrieval method, supporting semantic search, graph traversal, exact matching, and source-aware filtering.
- Artificial Intelligence
- Developer Tools
- AI Agents
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Long-horizon AI agents increasingly need to answer complex, multi-step questions that span time and relationships—yet most AI systems rely on single-mode retrieval that fails when queries require temporal awareness, graph traversal, or exact lexical recall. Standard document Q&A and vector-only search cannot reliably support the persistent memory that agents need for evolving knowledge bases, support histories, and research workflows.
Solution
Deyta is an engineering-first company building the data layer for reliable long-horizon AI systems. Its flagship product, Khora, is an open-source Python library that creates knowledge repositories from unstructured and structured multi-source data, exposing a single query substrate designed for integration into long-running AI agents. Khora ingests raw data into graph, vector, and relational storage, then uses its default VectorCypher retrieval engine to infer the best retrieval path for each query—combining semantic search, graph traversal, keyword matching, exact lookup, and source-aware filtering as needed. The system includes a temporal-aware memory layer with triple timestamps and forgetting-curve decay, plus an optional data consolidation phase that deduplicates entities, compacts repeated facts, and improves retrieval quality as the repository grows.
Target Audience
Primary users are AI engineers and developers building long-horizon agents for research assistance, enterprise knowledge graphs, and evolving knowledge bases that require persistent, reliable memory across unstructured and structured data sources.
Features
- Three storage types (graph, vector, relational) structured at ingestion time, enabling semantic, relationship, and exact-match queries from one system
- Query-aware hybrid retrieval via VectorCypher engine that routes to the appropriate method based on the query type
- Temporal-aware memory with triple timestamps, recency-aware scoring, and forgetting-curve decay for "what was true then, what changed" questions
- Data consolidation phase that deduplicates entities, clusters related events, updates relationship quality, and identifies stale or orphaned records
- Python library design for direct embedding into agent workflows
- Integrations with production infrastructure including PostgreSQL with pgvector, Neo4j, and agent frameworks such as CrewAI, LangGraph, Google ADK, OpenAI Agents SDK, and LlamaIndex