Knowledge

The knowledge graph is the firm's structured map of a tenant's product that is developed by firmd agents. It holds entities (e.g. customer segments, modules, features, dependencies), the relationships between them.

It is one component of institutional memory — not a synonym for all memory. Communication transcripts, content items, and telemetry events sit alongside it. The graph is what makes the rest of memory navigable.

Agents reasoning from raw prompt context invent plausible-but-wrong facts. The classic mitigation — paste more documents into the prompt — does not scale and erases the structure the documents originally had.

The knowledge graph fixes both ends. It gives agents a grounded reference they can browse and cite, resisting plausible invention. It also turns the slow human labour of building an ontology into something an agent can extend — pulling entities from briefs, proposing relationships, summarising narrative material — at the speed of product work.

This is also where the research converges. A 2025 ACL systematic review and a 2024 NAACL survey both find that integrating knowledge graphs with LLM inference measurably reduces hallucination and improves benchmark performance. The benefits span three families: knowledge-aware inference (the graph informs each model call), knowledge-aware learning (training against graph-structured supervision), and knowledge-aware validation (the graph checks whether outputs are claim-consistent). Industry results mirror the finding on the retrieval side — graph-based retrieval outperforms vector-only retrieval on relational and multi-hop questions, while vector retrieval remains faster and simpler for plain semantic lookup.

The honest limitation in the same reviews: the field has not converged on a single dominant integration pattern. firmd's bet on a knowledge graph is informed by this literature, not settled by it — the substrate is the right one; the precise integration mechanism is something we expect to keep iterating on.