Public beta, live and self-serve

Clean up the mess in AI-generated documents.

An LLM grading its own work is not a check. PortMem is the independent layer that verifies every claim from LLM-generated content against the documents you used to build it: what's wrong gets flagged, what's missing gets named, and a human signs off before it ships, with every decision securely recorded for future audits.

Open the console
Precision

It doesn't cry wolf. Two false alarms in 1,478 clean claims checked against declared ground truth across four regulated domains (99.9%). When PortMem flags something, it's worth your time.

Evidence

Every verdict arrives with its receipts. Each claim is checked against the governing version of your documents and cited to the exact passage, so a reviewer confirms in one glance instead of re-reading the source.

Proof

Every step on the record. Verdicts and sign-offs land in a signed, tamper-evident ledger your auditor can verify.

How it works

Submit a report. Get the audit.

1

Upload the governing documents; versions and amendments handled.

2

Submit the AI-written report, from any tool.

3

Get the audit: per-claim verdicts, evidence quotes, omissions, human sign-off.

4

Export the proof: a signed evidence packet an examiner can verify offline.

PortMem review console: delivery gate clear, contradicted claims with evidence and human sign-offs

The production console: real verdicts on public deal documents, verified by hand against the deed text. Agents can run the whole loop programmatically (MCP). Open the console →

The proof

Three regulated benchmarks. One pattern.

In all three, the right answer is not the most similar document. PortMem finds the one that governs.

← swipe to scroll →
Benchmark Question PortMem Same model, no documents Why it is hard
Finance · primary Which version is in force? 100% 95% 95% still means acting on a superseded standard one time in twenty. PortMem returns the controlling version with its citation, every time.
Legal Has this precedent been overruled? 99% 48% Overruling cases look nothing like the cases they replace. PortMem finds them anyway.
Pharma Has this drug been recalled? 99% 29% Standard retrieval finds the approval. PortMem surfaces the recall.

FASB ASC supersessions n=20, SCOTUS overruling pairs n=100, FDA recall lookups n=100. "No documents" = the model answering from its own knowledge (July 2026 run); PortMem column from May 2026 runs, same model and judge. Given the entire corpus pasted as context instead, the same model scores 100% / 96% / 80%, so PortMem's margin is largest where the corpus cannot be hand-fed. Full artifacts, raw outputs, scorer, and correction history: portmem.com/evidence.

Across four regulated domains and 255 seeded-error checks (pre-registered protocol, raw runs published): 92.9% of planted errors that entered review were caught or flagged, with two false alarms in 1,478 clean claims. Every miss is published, and corrected numbers supersede earlier ones in the open, on the evidence page.

Where PortMem fits

Generators grade their own homework. PortMem is the independent check.

← swipe to scroll →
Approach Up to date Audit Multi-document Model-agnostic Regulated review
Long-context LLMs
Sonnet 1M class
 ✗ ✗ ~ ✗ ✗
Naive RAG
LangChain, LlamaIndex
 ✗ ~ ✗ ✓ ✗
Agentic RAG
LangGraph, CrewAI, AutoGen
 ✗ ~ ✓ ✓ ✗
Pre-indexed document tools
Pinecone Nexus, PageIndex
 ✗ ~ ✓ ✓ ✗
Vertical legal AI
Harvey, CoCounsel, Hebbia
 ~ ✓ ~ ✗ legal only
Enterprise search
Glean, Sana
 ✗ ~ ✗ ✓ ✗
Vector store + rerank
Vectara, Pinecone, Cohere
 ~ ~ ~ ✓ ✗
PortMem
The verification layer
 ✓ ✓ ✓ ✓ ✓

Our assessment of each approach as of mid 2026, based on public documentation. Not independent benchmarks.

✓ Built in ~ Partial or workaround ✗ Not built
Research foundation

Built on four published arXiv papers: PhaseGraph · BridgeRAG · RegimeRouter · CAR. Each fixes a failure mode standard retrieval has in regulated content. Datasets & eval code →

Get started

The console is self-serve today.

Upload a document and run your first review in minutes. For a managed pilot, leave your email; finance teams first, legal, pharma, and security follow.

Open the console

Or write directly to andre@portmem.com.