Temporal Graph Memory connects your documents and returns original passages with sources and version history. No generative LLM in the retrieval engine.
Pilot access starts with a 30-minute call. We set up your account during onboarding. How access works
Scroll to exploreExample retrieval
One sample contract. Three ways to inspect the evidence your agent receives.
Illustrative sample data · not a live API call or a benchmark result
sample-contract.md v2
Invoices are issued monthly and are payable within thirty days of receipt.
Either party may terminate this agreement with sixty days' written notice.
↑ This exact passage is returnedSupport requests may be submitted by email on business days.
A fictional document created for this example.
Supporting passage found
“Either party may terminate this agreement with sixty days' written notice.”
The returned passage is the highlighted text in section 2. Your application can cite it directly.
sample-contract.md · section 2 · version 2
A source you can inspect, not just an answer to trust.
An exact quote still needs a relevance and completeness check. How source attribution works
Results from the documented evaluation, not a guarantee for every question or dataset. Answer quality is measured on answered questions. Engine timing excludes network overhead. Read the methodology and limitations →
One API call. Text and markdown in; other formats convert during onboarding.
Sentence-level passages that keep their exact bytes. No schema, no tagging.
Passages link by meaning. Search in ~22 ms, about half a second end-to-end.
Update or delete documents freely. Old versions keep answering exactly as they did.
Recorded results depend on the dataset and workload. Queries pause while a graph change indexes. Read about version history.
TGM cross-links every document set you give it, contracts, policies, research, reports, into one knowledge graph, so related passages surface together and every one still names its source.
TGM returns passages from your documents with a source identifier. Source fidelity is separate from whether a passage fully answers a question.
Responses include not_covered so your application can handle questions without returned evidence. Evaluation results are documented below.
Query a saved graph version to inspect the passages returned from that version of your documents.
The documented engine connects passages using embeddings. Your application can use the retrieved context with a language model.
RAG and GraphRAG capabilities depend on the implementation. We have not published a head-to-head evaluation here. Read the TGM evaluation.
Pilot access
Use the 30-minute call to discuss your documents, the questions your application needs to answer, and whether a pilot is a fit. Booking requests require confirmation.
Bring a description of the documents and a few representative questions. You do not need to upload documents to book.
Confirm scope, access, pricing, and how results will be evaluated before proceeding.
Your account and API key are set up during onboarding. Then connect the SDK and evaluate on your agreed document set.
Developers building AI applications that need document passages with source identifiers and version history. The current onboarding flow is a concierge pilot.
The retrieval engine returns original passages. Your application decides how to present them or use them with a language model. A downstream generated answer needs its own evaluation.
The documented ingestion flow accepts plain text and Markdown. Other formats, including PDF, are converted with you during onboarding.
You need a pilot account and API key first. Book a call to discuss fit and onboarding. Installing the package alone does not grant API access.
Queries pause while a graph change indexes. The pause depends on the size of the change. Saved versions remain part of the graph history.
152 of 161 answered questions were graded excellent in the documented evaluation. It does not describe coverage of every authored question or guarantee results on your documents.
For data-handling terms, read the privacy policy and terms of service. Confirm requirements for your documents during the pilot discussion.
We build Temporal Graph Memory for document retrieval in AI applications. Explore the documentation and evaluation before deciding whether to start a pilot.