Semantica overview and setup for AI agent context graphs

Semantica is presented as open source, self-hosted infrastructure for AI agent memory and decision records. The speaker describes an MIT-licensed Python package that sits beneath an existing LLM, vector store and agent framework. Its context graph stores entities, relationships and decisions as queryable nodes, with source provenance attached to facts.

The overview covers data ingestion, entity extraction, deduplication and conflict handling before explaining rule-based inference through Rete, Datalog and SPARQL. According to the speaker, graph construction, logical inference and lineage tracing do not require an LLM. Decision records include context, reasoning, outcomes and confidence; causal links support queries about precedents and downstream effects. The video also discusses W3C PROV-O exports, SHACL validation and historical snapshots.

Storage options include embedded Oxigraph and external graph or vector databases. Databricks and Snowflake connectors receive separate attention. The speaker reports node search improving from 24 milliseconds to 4 microseconds on a 118,000-node graph; this is a reported benchmark, not a general performance guarantee.

The quick start uses pip installation followed by Semantica doctor, with Python 3.8 or later. The speaker says the initial setup works locally without a server, while individual backends require additional modules. The overview also covers a browser graph explorer, MCP access, a REST API and CLI.