FIRME: Registry-Anchored Memory Graphs for Evolving, Heterogeneous Investor Knowledge

Yunbai Zhang, Yunyi Gao, Han-Tai ShiaoarXiv 2026 · arXiv

TL;DR

FIRME (Fusion and Identity Registry for Memory Evolution) is a framework for building persistent, evolving knowledge graphs that track investor interests across time and diverse data sources. By anchoring relationship histories to a registry of canonical identities and learned aliases, it prevents the fragmentation of evidence and provides an auditable, lifecycle-based view of how investor concerns wax and wane.

What Problem Does This Paper Address?

In the world of financial AI, understanding an investor requires looking at both what they say (unstructured text like call transcripts) and what they do (structured data like portfolio cash flows). However, building a coherent "memory" of these interactions is difficult for two main reasons:

  1. Identity Fragmentation: An investor might discuss "muni bonds" in one meeting and "municipal bonds" in another. Without a way to reconcile these different "surface forms" to a single identity, the AI creates separate, disconnected records, losing the cumulative history of that interest.
  2. Temporal Evolution: Interests aren't static. A concern about "market volatility" might spike during a crash, fade during a bull market, and resurface later. Standard knowledge graphs often struggle to represent this reinforcement, decay, and reactivation without losing the underlying evidence.

The authors address these gaps by introducing a registry-anchored memory graph that ensures new observations are correctly "reattached" to existing histories.

Methods

FIRME operates through a two-stage fusion process:

  • Short-Term Fusion: Within a single time period, FIRME combines text-extracted relationships with structured cash-flow signals. It uses a fixed product-topic ontology to link broad discussion topics (e.g., "Tax Efficiency") to specific financial products (e.g., a tax-exempt bond fund). If both modalities align, the relationship is labeled as "strong-interest."
  • Long-Term Fusion: This is the core innovation. FIRME maintains a Persistent Registry ($R^t$) of canonical identities and their known aliases. When new data arrives, the system attempts to match it using a tiered approach: first by exact ID, then by normalized name, then by known aliases, and finally via LLM-based semantic matching.
  • Memory Reinforcement-Decay: Relationships in the graph aren't just binary links; they have an "evidence strength" score. Recurring observations reinforce the score additively, while periods of inactivity cause the score to decay multiplicatively.
  • Lifecycle States: Based on these scores and match history, every relationship is assigned a state: Emerging, Reinforced, Stable, Fading, or Reactivated.
Figure 2: Memory-edge lifecycle with inactivity threshold 𝑁= 3. The entry arrow initializes a relationship as Emerg- ing. Solid transitions between states, including self-loops, denote matches; dashed transitions denote misses. Guards count consecutive misses including the current period; a match resets the count.

Datasets

The authors evaluated FIRME on two distinct corpora:

  1. Proprietary Investor-Call Corpus: A large-scale dataset from The Vanguard Group consisting of 10 months of investor call transcripts paired with monthly portfolio cash-flow records (~1,200 calls per month).
  2. S&P 500 Earnings Transcripts: A public dataset used to test the system's ability to handle a broader entity schema (companies, geographic regions, financial metrics) across three reporting periods.

Key Insights and Ideas

  • Aliases are the "Glue": The most significant finding is that simply having a canonical registry isn't enough. Storing and reusing aliases (alternative names used in the past) is crucial for high-recall reattachment. It allows the system to resolve future mentions deterministically, saving both computation time and LLM tokens.
  • Consolidation vs. Fragmentation: Without FIRME, the graph becomes cluttered with near-duplicates. The system effectively collapses multiple surface variants into a single node, as shown in the comparison between standard LLM normalization and FIRME's registry-based approach.
Topic consolidation: FIRME (b) collapses 11 different surface variants into 5 canonical topics compared to standard LLM-Norm (a).
  • Auditable Trajectories: By using a deterministic state machine for lifecycles, the system provides an "explainable" memory. A human can look at a "Fading" interest and see exactly which periods of inactivity led to that status.

Key Results

The authors used a "Correct Match Rate" (CMR) to measure how often incoming observations were correctly linked to their historical counterparts.

  • Performance Gains: Full FIRME increased the final-period CMR by 7.8–12.1 percentage points compared to a control that lacked alias memory.
  • Human Audit: In a blinded human audit, FIRME showed reattachment-recall gains of 22.5–29.0 points. This means it is significantly better at "remembering" that a new mention belongs to an existing relationship.
  • Efficiency: By resolving recurring aliases deterministically, FIRME reduced semantic-matching token consumption by 40.3% compared to the no-alias control.

Limitations and Future Work

While FIRME represents a step forward in agentic memory, the authors note several limitations:

  • Deterministic Logic: The reinforcement and decay parameters are currently fixed. Future versions could learn entity-specific decay rates (e.g., an interest in "Retirement Planning" might decay slower than an interest in a specific "Meme Stock").
  • Negative Signals: The current model focuses on positive evidence (discussions and inflows). It does not yet explicitly model negative signals like outflows or expressed dissatisfaction.
  • Ontology Dependence: The short-term fusion relies on a manually curated product-topic ontology, which may be difficult to scale or adapt to rapidly changing financial markets.

Future work aims to explore firm-level aggregation and testing the framework over much longer temporal windows to capture multi-year seasonality.

Summary updated

Abstract

Financial AI workflows that reuse investor knowledge need persistent memory that integrates observations across sources and periods without losing identity or provenance. In a knowledge graph, recurring observations may use different surface forms, fragmenting relationship histories and preventing evidence from accumulating. We introduce Fusion and Identity Registry for Memory Evolution (FIRME), a registry-anchored memory graph that reconciles incoming observations to persistent canonical relationships and retains their source, time, and context. Within periods, rule-based fusion links investor-call discussions to portfolio cash-flow signals, directly or through a fixed product-topic ontology. Across periods, canonical identities and persistent aliases direct recurring observations to the same relationship history; deterministic updates summarize reinforcement, inactivity, and reactivation. We evaluate FIRME on proprietary investor-call and cash-flow records and on S&P 500 earnings-call transcripts. Compared with a historical registry that retains canonical identities and semantic fallback but omits alias memory, FIRME increases final-period Correct Match Rate by 7.8-12.1 percentage points across four dataset-backend settings under a fixed-judge and fallback-review protocol. A blinded human audit of 200 incoming observations per setting finds Correct Match Rate gains of 9.0-16.5 points and reattachment-recall gains of 22.5-29.0 points, with higher observed merge precision in all four samples. A separate audit finds valid evidence paths for 293 of 300 outputs from full within-period fusion in the investor-call setting. Extraction-faithfulness and round-trip diagnostics assess source grounding and preservation of extractor-derived graph content.
BibTeX
@misc{zhang2026firme,
  author = {Yunbai Zhang and Yunyi Gao and Han-Tai Shiao},
  title = {{FIRME: Registry-Anchored Memory Graphs for Evolving, Heterogeneous Investor Knowledge}},
  year = {2026},
  eprint = {2610.40482},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.40482},
}
Figure