Back to AI Research

AI Research

FIRME preserves relationship identity and evidence across changing financial records

Key Takeaways

  • A registry-backed memory graph reduces fragmented matches in the paper’s evaluated investor-knowledge datasets.
  • ### A repeated relationship should keep its history
  • The [FIRME paper](https://arxiv.org/abs/2610.40482) addresses a memory problem in financial AI workflows: incoming records may describe the same relationship using different names or wording.
  • If each version becomes a separate graph entry, supporting evidence becomes fragmented across sources and time.
  • FIRME uses a persistent registry of canonical relationships and aliases.

A repeated relationship should keep its history

The FIRME paper addresses a memory problem in financial AI workflows: incoming records may describe the same relationship using different names or wording. If each version becomes a separate graph entry, supporting evidence becomes fragmented across sources and time.
FIRME uses a persistent registry of canonical relationships and aliases. New observations attach to the corresponding relationship history while retaining their source, time and context. The aim is to accumulate evidence without losing the identity of what the evidence concerns.

Combine sources without losing provenance

Within a period, rule-based fusion connects investor-call discussions with portfolio cash-flow signals, directly or through a fixed product-topic ontology. Across periods, the registry and aliases help recurring observations return to the same history. Deterministic updates summarize reinforcement, inactivity and reactivation.
These are mechanisms for organizing knowledge. A remembered relationship or a valid graph path does not itself establish that an investment conclusion is correct, that the underlying record is complete or that a portfolio will perform better.

The comparison isolates alias memory

The evaluation uses proprietary investor-call and cash-flow data alongside S&P 500 earnings-call transcripts. Compared with a historical registry that retains canonical identity and semantic fallback but omits alias memory, FIRME improves final-period Correct Match Rate by 7.8 to 12.1 percentage points across four dataset-backend settings.
A separate blinded human audit reports matching and reattachment gains. Another audit finds valid evidence paths for 293 of 300 outputs in the investor-call fusion setting. Those counts describe the tested memory and provenance tasks, not financial returns.
For an applied evaluation, inspect a recurring relationship that disappears and later returns under a different alias. Check both whether it reconnects and whether the retained evidence still points to the right source. Matching more often is useful only when mistaken merges and provenance remain visible too.

Comments