Knowledge graphs for trading: memory needs time, provenance and tests

The reader question is whether a knowledge graph makes an AI trading agent more useful. Our answer is conditional: it may help retrieve connected research and explain how a conclusion developed, but the storage label does not establish that retrieved information is correct or chronologically admissible. Begin with the questions the memory must answer, then compare implementations.

A useful record distinguishes a source observation, an analyst’s interpretation, a strategy hypothesis and an eventual outcome. Linking them preserves a research trail. Merging them into one free-text “fact” risks turning a tentative explanation into a durable premise. This article proposes an evaluation framework; it does not benchmark VC Trader AI or claim that graphs outperform other stores.

Economic releases make the timing problem concrete. The St. Louis Fed explains that published series can change as errors are corrected, methods evolve or more complete observations arrive. ALFRED preserves vintages rather than only today’s view of the series. [1]

A backdated observation date is therefore insufficient for reconstructing a decision. Record at least the period described, the source’s release time when available, the ingestion time and the source version. For intraday analysis, a date-only archive may still be too coarse. The acquisition process must establish what was available before the decision timestamp.

The common conceptual distinction is between valid time, when an assertion concerns the world, and knowledge time, when the system learned it. In our proposed trading record, these are explicit fields rather than assumptions inferred from a document title. If the original release time cannot be established, mark that uncertainty and exclude the item from claims requiring intraday precision.

A graph can then link a research claim to its source, affected instruments, assumptions and tests. A relational database with versioned records could also implement much of this contract. The schema is valuable because it makes the acceptance criteria explicit before choosing a database.

W3C’s PROV family describes how to represent the entities and activities involved in producing information. It is a useful reference for recording where a research artifact came from. It is not a truth validator or a financial certification. [2]

For a graph relationship such as “company exposed to currency,” preserve the evidence and scope: a filing’s reporting period, the business segment concerned and whether the relationship is observed or inferred. A path connecting two companies through a shared supplier is a research lead. It does not establish the direction or magnitude of a tradable price effect.

Corrections should append or supersede an earlier assertion with a reason, keeping its historical state inspectable. A failed hypothesis can remain useful memory if it is clearly marked. Deleting every unsuccessful conclusion would make the archive look more confident than the research process actually was.

LangGraph’s documentation separates thread-level checkpoints from stores for information across threads. It also notes that in-memory checkpoints disappear after a process restart. These are implementation details of that framework, not evidence about another platform. [3]

Retrieval-augmented generation supplies external material to a model; the foundational RAG paper evaluates that arrangement on NLP tasks. GraphRAG adds a different technique: entities and community summaries to help answer broad questions across document collections. Neither source establishes point-in-time correctness for market decisions. [4] [5]

Ask which problem needs solving. Resuming a paused investigation, finding one original release, summarising recurring themes and reconstructing historical knowledge are different tasks. A visually sophisticated network can leave the hardest one unresolved.

The September 2026 META paper studies episodic memory for financial decision-making. Its method retrieves historical market episodes; its experimental setup uses four-hour OHLCV sequences and historical validation windows. The authors’ arXiv record reports acceptance in AACL-IJCNLP 2026 Findings. We have not reproduced the evaluation. [6]

That is relevant evidence that memory design is an active research question. It should not be presented as a live high-frequency execution benchmark or a general finding that graph memory earns alpha. The distinction matters because a mechanism can be promising while its experimental scope remains narrower than the language used to describe it.

The older “Lost in the Middle” study also found position-dependent retrieval failures in the models and tasks it evaluated. It motivates testing whether information is actually used, rather than assuming that adding a long context solves retrieval. It does not supply a failure rate for today’s models. [7]

We propose six acceptance cases. First, ask for an original release while a later revision is present. Second, request a historical answer whose correct current answer differs. Third, introduce two entities with similar names. Fourth, supply conflicting sources and require the conflict to remain visible. Fifth, restart the process and test the promised persistence. Sixth, query from a workspace that lacks permission to see the record.

For each case, retain the expected evidence, retrieved records and final answer. Score citation correctness separately from answer correctness. A response might quote the right number from the wrong vintage, or reach a plausible answer without admissible evidence. Report those as distinct failures. Include an abstention case where no accessible source supports a conclusion.

Compare a versioned-document baseline with the proposed graph on identical inputs and tasks. Hold extraction rules and access scope constant where possible. Measure retrieval quality, source traceability, freshness violations, latency and total processing cost. A useful comparison must count ingestion and correction work as well as query time. This is a test design, not a completed experiment.

Graphs are attractive when relations themselves are central to the question. Versioned documents or structured tables may suffice when the main task is finding a specific release. Either choice can fail through bad extraction, stale records or incorrect access control. Revise the architecture when a controlled comparison shows a material advantage for the workload.

VC Trader AI’s public copy describes persistent workspace context and an intelligence graph connecting research, instruments, decisions and outcomes. That is a first-party scope statement. This article’s proposed temporal tests are evaluation requirements, not a claim that every field or test is already implemented. Review the market-intelligence description with those boundaries in mind.

The practical conclusion is to demand one reconstructable research decision. Retrieve its original sources, reproduce the knowledge available at the time and show what changed afterwards. If a memory system cannot do that, additional relationships may increase the amount stored without increasing what can be trusted.