Vektori

Vektori is an open-source memory engine designed for AI agents, focusing on structured memory that tracks not only facts but also the context and evolution of information. Its core architecture is a three-layer sentence graph, which organizes memory into Facts (L0), Episodes (L1), and Sentences (L2). This structure enables agents to recall not just isolated facts, but also behavioral patterns and the trajectory of conversations, supporting deeper temporal and contextual understanding.

The platform allows users to ingest chat data and retrieve memory at varying levels of depth. At the L0 layer, agents can access distilled facts for fast planning and lightweight operations. The L1 layer adds episodes that summarize behavior trends and shifts across sessions, while L2 provides full traceability back to the original source sentences, enabling complete trajectory replay. Each memory item is grounded in source conversation evidence, and the sentence sequencing preserves information about what changed and when. Vektori supports pattern discovery and helps agents understand not just preferences, but also how and why those preferences have changed over time.

Vektori is delivered as a Python-first API and can be installed via pip. It defaults to using SQLite for local development without requiring Docker, but also supports production-ready backends such as Postgres with pgvector, as well as graph and vector databases like Neo4j, Qdrant, and Milvus. There is also an in-memory mode suitable for CI, tests, and reproducible evaluations. The engine is compatible with various embedding and extraction models, including those from OpenAI, Azure, Anthropic, NVIDIA, LiteLLM, and local model paths, allowing integration with a preferred model stack without rewriting agent loops.

0 open-source license and is intended for developers building AI agents that require persistent, structured memory and contextual retrieval. Benchmarks indicate high retrieval accuracy on conversational memory tasks, and the project actively publishes evaluation results and welcomes community contributions.