Persistent, traceable reasoning for AI systems in finance, legal, and compliance. Multi-hop retrieval with full decision provenance.
Multi-hop retrieval benchmarks · Research evaluation results
Standard RAG systems treat every query as a fresh start. They retrieve, answer, and forget. Enterprise decisions require more.
When an AI system makes a recommendation, there is no way to trace which documents informed it, whether any were superseded, or why earlier evidence was discarded.
Complex questions require reasoning across multiple documents — "What regulation superseded the rule that this contract references?" Standard vector search collapses on these.
In legal and regulatory documents, what matters is the controlling document — the one that is currently active. Semantic similarity retrieves the wrong answer confidently.
Inspired by how divergent thinking works — capturing broadly, consolidating deliberately, retrieving with awareness of authority and context.
Every document is ingested without premature filtering. Content is embedded, described in context, and tagged — preserving raw information before structure is forced on it.
Background workers extract knowledge into a graph — deduplicating entities, discovering cross-domain bridges, and detecting contradictions between episodic memories.
Queries trigger two-phase retrieval: vector search fused with graph traversal via Personalized PageRank — surfacing controlling documents, not just similar ones.
Every feature is designed for high-stakes retrieval where wrong answers have real consequences.
Every retrieval result carries full provenance — which memories informed it, when they were captured, and whether any contradictions were detected along the way.
Answers complex questions that require chaining across multiple documents. Graph-based traversal discovers paths that pure embedding search cannot see.
In authority-governed corpora (legal, regulatory), retrieves the active controlling document — not the most similar one. Knows when regulations are superseded.
The system learns domain interests over time. Retrieval results are weighted by relevance to tracked expert domains, giving analysts the context they actually need.
Route different tasks to different models. Run fully local with Ollama at zero cost, or use cloud models for critical extraction. Cost profiles from $0 to $80/mo.
The knowledge graph maintains CONTRADICTS edges. When a retrieved answer is in tension with stored knowledge, it surfaces the conflict — not just the most recent answer.
Designed for domains where the wrong document — or the outdated one — has real consequences.
SEC filings, regulatory compliance, FINRA rules
Case law, precedent tracking, statutory authority
CVE tracking, patch releases, GHSA advisories
FDA drug labels, clinical trial data, recalls
Early-stage public research — not peer-reviewed, but grounded in empirical benchmarks and real-world evaluation.
For authority-governed corpora, the retrieval target is the active frontier of the authority closure — not the highest-scoring document. Benchmarked across FinSuperQA, SCOTUS, GHSA, and FDA.
Original question embeddings outperform bridge passages for multi-hop retrieval. Bridge passages contribute via entity identity, not embedding similarity — a counterintuitive finding with major efficiency implications.
Systematic study of bridge passage retrieval strategies for complex multi-hop questions. Establishes baseline methods and failure modes that motivated Question-Space Dominance.
The production system combining all findings: neurodiversity-inspired memory architecture with episodic/semantic duality, interest tracking, and hybrid convergent/divergent retrieval.
Quality measured across 4 orthogonal axes
PortMem is in research preview. We are working with a small number of enterprise partners in finance and legal to validate the system in production environments.
No spam. We will reach out personally to discuss your use case.