Semantica is the #2 repository on GitHub Trending this week, up 5,135 stars in seven days and 1,181 today, now at 7,560 stars and 784 forks. It describes itself as "Graph-Native Infrastructure for Context and Accountable AI Systems," and the README's own subtitle is blunter: the open source Palantir for AI agents. It is MIT-licensed Python, installable with pip install semantica, and despite the sudden attention the repository has existed since 25 June 2025 — this is a project reaching an inflection point, not a fresh drop.

The framing is a compliance argument rather than a performance one. Most agents store embeddings, not meaning: context that cannot be explained, decisions that cannot be audited. In a regulated domain that is an exposure, because an underwriting agent's approval has to survive a regulator's "why" months after the fact. Semantica's response is to sit underneath your LLM, vector store, and agent framework as a deterministic layer — and the claim worth highlighting is that no LLM is required for graph construction, reasoning, or provenance. That is what makes its answers reproducible.

The headline abstraction is the Context Graph, where every entity, relationship, decision, and fact is a first-class queryable node. A decision is not a log line: record_decision() writes a structured node with scenario, reasoning, outcome, and confidence; add_causal_relationship() links it upstream and downstream using one of CAUSED, INFLUENCED, or PRECEDENT_FOR; trace_decision_chain() walks the full causal ancestry; find_similar_decisions() does semantic precedent search across everything recorded; analyze_decision_impact() maps downstream influence; and check_decision_rules() acts as a policy gate. The whole trail exports as W3C PROV-O, plus CSV or JSON, in the format compliance frameworks accept for regulator submission.

Underneath that sits a full pipeline, and each stage is an independently importable module: ingest (files, web, databases, streams via Kafka and Kinesis, Git, email, MCP, plus native Databricks Unity Catalog and Snowflake connectors), split for GraphRAG-native entity-aware chunking, semantic_extract for NER, relations, events, and triplets, then conflict detection and deduplication — conflicting facts get flagged rather than silently overwritten, which is exactly where a vector index fails. Reasoning covers forward chaining, a Rete network, Datalog, and SPARQL with explainable paths; governance covers SHACL constraints, OWL generation, and SKOS vocabularies. Storage is deliberately polyglot: RDF triple stores (embedded Oxigraph, Blazegraph, Apache Jena, Eclipse RDF4J) and labeled property graphs (Neo4j, FalkorDB, Apache AGE, AWS Neptune) alongside FAISS, Qdrant, Weaviate, Milvus, Pinecone, and PgVector — swappable without touching your code. Access arrives through a REST API, an MCP server, a CLI, and native Agno multi-agent support, with semantica doctor verifying an install. Bi-temporal facts and state_at() point-in-time snapshots let you replay the graph as it existed on a past date without reprocessing.

The honest counterweight is scope. This is a large surface — fifteen-plus modules, four reasoning engines, ten-plus graph and vector backends — and adopting it means committing to ontology and provenance discipline your team may not currently practise. It is squarely aimed at finance, healthcare, legal, government, and defense, so a hobby agent inherits the ceremony without the regulatory payoff. The version line is still pre-1.0 at v0.6.5 (11 August 2026) with 60 open issues, and a project that added two-thirds of its stars in one week has not yet been stress-tested by the audience that just arrived.