Semantica wants agent memory you can audit
Semantica went from nowhere to +967 stars in a day, sitting near the top of GitHub trending with 4,000 total. MIT licensed, Python, and the pitch is blunt: most agents store embeddings, not context, and an embedding cannot testify.
What it actually is, is a deterministic layer that sits under your model and your vector store and builds an actual graph. Every decision the system makes becomes a first-class node with W3C PROV-O provenance attached, so you can walk backwards from an output to the facts and rules that produced it. Reasoning runs on forward chaining, Rete networks, Datalog and SPARQL — no LLM in the loop for the inference itself, which is the whole point, because an explanation generated by a language model is not an explanation. It does bi-temporal fact tracking and point-in-time snapshots, detects contradictory facts before they poison the graph, speaks both RDF triple stores and property graphs like Neo4j, and ships native Databricks and Snowflake connectors.
Read the connector list and the target market is obvious: finance, healthcare, legal, government, defense. People who will eventually have to explain an agent's decision to a regulator and cannot answer "the vectors were close."
The bet underneath is that the memory layer splits in two. Fuzzy recall stays in embeddings because that's what they're good at. Anything load-bearing — what did the system know, when did it know it, why did it act — moves into something structured and queryable. Everyone building agent memory this year has been discovering the same thing the slow way. This one starts there.
https://github.com/semantica-agi/semantica
← Back to all articles
What it actually is, is a deterministic layer that sits under your model and your vector store and builds an actual graph. Every decision the system makes becomes a first-class node with W3C PROV-O provenance attached, so you can walk backwards from an output to the facts and rules that produced it. Reasoning runs on forward chaining, Rete networks, Datalog and SPARQL — no LLM in the loop for the inference itself, which is the whole point, because an explanation generated by a language model is not an explanation. It does bi-temporal fact tracking and point-in-time snapshots, detects contradictory facts before they poison the graph, speaks both RDF triple stores and property graphs like Neo4j, and ships native Databricks and Snowflake connectors.
Read the connector list and the target market is obvious: finance, healthcare, legal, government, defense. People who will eventually have to explain an agent's decision to a regulator and cannot answer "the vectors were close."
The bet underneath is that the memory layer splits in two. Fuzzy recall stays in embeddings because that's what they're good at. Anything load-bearing — what did the system know, when did it know it, why did it act — moves into something structured and queryable. Everyone building agent memory this year has been discovering the same thing the slow way. This one starts there.
https://github.com/semantica-agi/semantica
Comments