Semantica — the graph is for the audit, not the recall

Semantica (MIT, Python, 10,659 stars, first commit June 2025) calls itself the open-source Palantir for AI agents. It ingests enterprise data, extracts entities into a context graph and a knowledge graph, and runs analytics and causal reasoning over both — with decision provenance recorded throughout. Polyglot storage, RDF and LPG, W3C standards, self-hostable.

The framing sentence is the useful one: most AI agents act without a trail; they store embeddings, not meaning. Their example is underwriting — an approval has to survive a regulator asking “why” months later, and a cosine similarity is not an answer to that question.

Notably, the reasoning layer is deterministic and needs no LLM. The graph is infrastructure underneath the model, not another thing the model is asked to maintain.

The idea worth extracting

For a solo builder the platform is oversized — this is aimed at regulated industries with compliance exposure, and adopting it would be adopting an ontology practice, not a library. But the distinction it draws is real and transferable:

A retrieval graph and a provenance graph answer different questions. Retrieval optimises for finding the right thing now — which is what project-solograph’s code graph and vector indexes do, and what the tiering borrowed in graphify-vs-solograph makes cheaper. Provenance optimises for reconstructing why something happened, later — a different shape, different write path, different cost.

Conflating them is the common mistake, and it usually shows up as a retrieval store quietly failing at forensics: you can find the document the agent read, but not establish that this is why it did what it did.

decision-traces-compound argues the same case from the value side — that traces are the asset that accumulates. Semantica is what it looks like when someone builds the infrastructure for that claim rather than asserting it. And claude-code-anatomy’s finding that 98.4% of a harness is operational infrastructure predicts this: provenance is infrastructure, so it does not arrive by being a good idea.

Cheap version for one person: the session log already is a provenance store if nothing model-visible bypasses it. That is the same invariant deepseek-harness enforces at runtime — model-visible means logged. A graph on top is an optimisation for querying it, not the thing that makes provenance exist.

Links

• context-graphs-summary — the concept this productises
• decision-traces-compound — why traces are the compounding asset
• project-solograph — our graph, built for retrieval; the contrast is the point
• graphify-vs-solograph — where tiered loading was already borrowed from a neighbour in this space
• deepseek-harness — the cheap invariant that gives provenance without a graph
• agent-toolkit-landscape — belongs in the observability/governance rows