Automated Financial Research

The problem

Investment research is a race against time and volume. When a new filing or earnings release drops, analysts have to read the document, listen to any accompanying call, cross-reference previous statements, check sentiment, and update the underlying models — all before anyone else publishes. Human analysts are bottlenecked by the sheer volume of material, and traditional keyword search helps find documents, not answers.

What the workflow actually needs is a system capable of performing deep, multi-document synthesis autonomously and producing a draft note that a senior analyst can verify and publish. The output does not have to replace the analyst; it has to collapse the hours of collection work that precedes the analyst's real judgement.

Why these patterns

Agentic RAG is the engine that prevents hallucination in high-stakes analysis. Asked how a management team's stance on a topic shifted between two reporting periods, a standard RAG system would retrieve paragraphs containing the relevant keywords and hope for the best. Agentic RAG works differently: it retrieves the earlier document, extracts specific claims, retrieves the later one, compares them, and if the evidence is incomplete it issues a new query — for example, for news during the intervening period. It plans, retrieves, validates, and refines rather than generating in one shot.

Tripartite cognitive memory is what makes the agent behave like a firm's own analyst rather than a generic LLM.

• Semantic memory connects the agent to the firm's broader knowledge graph — the relationships between entities, sectors, and historical events.
• Procedural memory enforces voice, formatting, and compliance rules, so outputs consistently look and read the way the firm expects.
• Working memory keeps the agent focused on the specific subject and period it is analysing, ensuring facts from a parallel task do not bleed in.

The agent-native lakebase provides the foundation underneath both. Rather than forcing the agent to query a traditional relational store (which struggles with text) or a pure vector store (which struggles with tabular data), the agent-native architecture offers federated, multimodal access. The same agent can reach semantic embeddings for qualitative intent and structured queries for historical numeric data inside a single reasoning step.

What breaks without tripartite memory

Without isolating cognitive memory types, the system drifts toward two linked failure modes: compliance drift and context bleed.

When semantic facts and procedural rules are mixed into the same prompt, the model tends to prioritise the rich semantic material — the latest news, the punchy narrative — over the quieter procedural rules such as standard risk disclosures or restricted language. The resulting output might read well but fail review, which in a regulated setting translates directly into exposure.

Separating procedure (how to write) from semantics (what to know) keeps the compliance scaffolding rigid while leaving the insight layer dynamic. The agent can be creative with the thesis and strict with the template at the same time.

Operational considerations

Running research pipelines in a regulated environment is an exercise in controls.

Enforce citations. Reviewers cannot trust output they cannot verify. The agentic RAG pipeline must append block-level citations to every quantitative claim and every specific qualitative observation, so any reader can jump to the source in one click.

Data provenance is critical. The lakebase has to track when each document was ingested and which version of the embedding and extraction models processed it. When an error appears in a draft, the operations team needs to trace whether the issue came from agent logic or from stale data underneath.

The human is the editor, not the writer. The goal is not to remove analysts; it is to raise the floor of what they start with. The pipeline should produce a highly accurate draft, and the senior analyst's job shifts from collection to verifying the thesis, adding intuition the data cannot supply, and approving publication.