PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yunbai Zhang, Yunyi Gao, Han-Tai Shiao

arXiv:2610.40482v1cs.AIcs.CL

TL;DR

Heterogeneous sources and changing names can split observations of the same relationship, limiting accumulation of a persistent, provenance-preserving evidence history. FIRME fuses structured and unstructured evidence within periods and uses canonical identities plus persistent aliases to route recurring observations into relationship histories across periods. FIRME increased CMR over the no-alias control by 7.8–12.1 percentage points; human-audit reattachment-recall gains over that control were 22.5–29.0 points.

  • Problem

    Heterogeneous sources and changing names can split observations of the same relationship, limiting accumulation of a persistent, provenance-preserving evidence history.

    Source: §1 · p. 1

  • Method

    FIRME fuses structured and unstructured evidence within periods and uses canonical identities plus persistent aliases to route recurring observations into relationship histories across periods.

    Source: §1 · p. 1

  • Results

    FIRME increased CMR over the no-alias control by 7.8–12.1 percentage points; human-audit reattachment-recall gains over that control were 22.5–29.0 points.

    Source: §6 · p. 9

  • Takeaways & Limitations

    Correct registry reattachments extend one cumulative evidence history, while inactivity lowers evidence strength without removing earlier evidence.

    Source: §1 · p. 1, §5 · p. 9

    The proprietary investor-call and ontology resources are unavailable, the S&P corpus lacks paired cash-flow records, and neither window spans an annual cycle.

    Source: §7 · p. 10

Abstract

from arXiv · show

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.

1 Introduction

Investor knowledge must combine discussion and behavior while tracking changing relationships without losing identity or evidence. FIRME couples within-period fusion with registry-anchored updates to persistent relationship histories.

1.2 Contributions

  • FIRME’s stable head and reconciled target identities direct observations to persistent histories, which retain each observation’s relation, source, time, and context.Matched observations reinforce a history; unmatched periods lower its evidence strength while earlier evidence remains available.Source: §1 · p. 1, §1.2 · p. 2

1.1 Related Work

  • Prior graph work addresses construction-time canonicalization, temporal validity or retrieval, and financial KG integration; FIRME focuses on recurring investor relationships across heterogeneous sources and periods.Investor interests may recede and resurface, while conversational and structured evidence can corroborate a relationship without being duplicates.Source: §1.1 · p. 2

1.2 Contributions

  • FIRME improves final-period CMR by 7.8–12.1 percentage points across four settings; blinded audits find 22.5–29.0-point reattachment-recall gains and higher merge precision in all four samples.The comparison is against a no-alias control retaining historical canonical reconciliation and semantic fallback.Source: §1.2 · p. 2
  • In the final-period investor-call audit, 293 of 300 full-fusion outputs had valid evidence paths, including 91 of 94 triggered by Rule 4 but not Rule 3.The derived product relationship references supporting observations without merging topic and product identities.Source: §1.2 · p. 2
  • Fixed lifecycle rules summarize relationship evidence trajectories using match status, prior state, and consecutive unmatched periods.Assigned states are traceable to retained observations and periods without a match.Source: §1.2 · p. 2

2 Problem Formulation

The formulation combines transcript-derived and structured financial graphs within a period, then reconciles canonical relationships across periods. It also addresses entity normalization so variants can share a relationship history.

2.1 Preliminaries and Notation

  • A knowledge graph represents directed, typed edges between entities; investor-call graphs use three entity types and relation types, while S&P 500 transcripts use a broader entity schema.The graph is written as G = (V, ℛ, E), with edges (u, r, v).Source: §2.1 · p. 3

2.2 Data Modalities

  • Transcript text and meeting notes yield an unstructured graph of person-or-organization relationships to products or topics identified as substantively discussed.This graph is denoted G_t^u.Source: §2.2 · p. 3
  • Structured records yield a graph from predefined columns and value conditions, with a primary index identifying the person or organization.A capital inflow into a product can instantiate an investor–product behavioral edge.Source: §2.2 · p. 3

2.3 Problem Statement

  • Short-term fusion combines the unstructured and structured graphs through a fixed product–topic ontology, assigning strong-interest only when both modalities corroborate a relationship.Strong-interest is a rule-based evidential label, not a calibrated preference or stance estimate.Source: §2.3 · p. 3
  • Long-term fusion stores canonical relationships as time-indexed edges whose attributes accumulate observations, supporting evidence, first- and last-seen periods, memory strength, and lifecycle state.The current short-term graph is normalized through the prior registry before reconciliation with prior memory.Source: §2.3 · p. 3
  • The registry and memory dynamics initialize empty; first-period candidates are deduplicated by lemmatized equality and then semantic matching, while later unresolved candidates are compared with the prior registry.At the first period, normalized edges enter memory in Emerging state with a baseline score of 0.5.Source: §2.3 · p. 3
  • Entity normalization determines whether text-extracted names refer to the same entity; without it, near-duplicates fragment relationship histories and burden both fusion stages.The registry-based solution is presented in Section 3.Source: §2.3 · p. 3

3 Methodology

FIRME constructs investor relationship graphs with schema-controlled extraction and within-period cross-modal fusion, then reconciles identities and updates relationship histories across periods. Its graph preserves observation provenance while tracking memory strength and lifecycle.

3.1 Entity and Relation Schema

  • The investor-call schema links person/org heads to products and topics through discussed, cf-signal, and rule-derived strong-interest relations; the actual speaker is recorded separately from the graph head.The S&P 500 transcript-only evaluation uses a broader entity schema and omits the investor-call cash-flow fusion rules.Source: §3.1 · p. 3

3.2 Knowledge Graph Construction

  • LLM2Graph preserves source head IDs and normalizes mentions locally; the persistent registry then resolves candidates through identifier, normalized-name, alias, and semantic matching.Structured cash-flow edges represent investor capital inflows to products and are retained only when monthly inflow exceeds $100K.Source: §3.2 · p. 4

3.3 Short-Term Fusion

  • Within periods, direct same-investor product corroboration or ontology-linked topic/product evidence assigns strong-interest to the product, while the topic remains discussed.The product–topic ontology is fixed before evaluation, and corroboration does not replace the original source indicators of observations.Source: §3.3 · p. 4

3.4 Long-Term Memory Fusion

  • Canonical target reconciliation checks identifiers, normalized names, and aliases before semantic matching; each persistent relationship receives one reinforcement or decay update per period.For unresolved entities, semantic matching selects the highest-scoring same-type candidate with a score of at least 0.8; matched edges gain 0.2, while unmatched edges retain 0.8 of prior strength.Source: §3.4 · p. 4
  • Each edge retains observation provenance and evolving strength; with N=3, it stays Stable through three misses, becomes Fading on the fourth, and remains stored.Scores decay from the first unmatched period, including while an edge is Stable; zero-score edges remain available for reactivation.Source: §3.4 · p. 5

4 Empirical Evaluation

FIRME is evaluated for identity continuity, source faithfulness, graph-content preservation, and evidence supporting within-period product corroboration. The evaluation combines two corpora, multiple backends, automated judgments, and human audits.

4.1 Evaluation Criteria and Setup

  • The proprietary investor-call corpus pairs transcripts with monthly cash-flow records across 10 periods; the S&P 500 corpus covers 297 companies across three reporting periods and tests extraction and temporal reconciliation, not cash-flow fusion.The proprietary records are unavailable publicly; PII was masked.Source: §4.1 · p. 5

4.2 Memory Usability

  • 7.8–12.1 percentage points: full FIRME raises final-period CMR over FIRME w/o Alias Memory across four settings under the judge and fallback-review protocol.The canonical-only control also exceeds LLM-Norm in every setting.Source: §4.2 · p. 6, Table 3 · p. 7
  • FIRME achieved the highest Edge Compression Rate and Entity Compression Rate in every setting; Figure 3 shows topic consolidation from 11 surface-form topics under LLM-Norm to five canonical topics while preserving distinct topics.The human audit sampled 200 final-period incoming relationship units per dataset–backend setting and used a shared reference across strategies.Source: Figure 3 · p. 6, §4.2 · p. 7
  • Human audit found 9.0–16.5-point CMR gains and 22.5–29.0-point reattachment-recall gains for full FIRME over FIRME w/o Alias Memory across four settings.Observed merge precision was higher in all four samples, ranging from 85.6% to 97.5%.Source: §4.2 · p. 7
  • For approximately 1,300 final-period investor calls using Opus 4.6, full FIRME reduced semantic-matching token consumption by 40.3% and total consumption across two stages by 7.2% versus FIRME w/o Alias Memory.Measurements exclude earlier registry-construction costs.Source: §4.2 · p. 8, Table 4 · p. 8

4.3 Extraction Faithfulness

  • Extraction-faithfulness diagnostics assess quotation verification, entity traceability, and normalization, but do not separately test relation entailment, direction, stance, or graph completeness.On investor calls, Opus 4.6 had lower hallucination and normalization error than GPT-5.4-mini; on S&P 500, backend results traded off across those metrics.Source: §4.3 · p. 8, Table 5 · p. 8

4.4 Round-Trip Recovery

  • Round-trip recovery found investor-call node- and edge-level F1 of 0.923–0.931 across backends, while S&P 500 recall reached 0.697 and preservation was lower and more backend-sensitive.Adjusted precision remained 0.952–0.982; lower recall reflected greater loss of reference-graph content.Source: §4.4 · p. 8, Table 6 · p. 9

4.5 Human Audit of Short-Term Fusion

  • In the final-period Opus 4.6 human audit, 293 of 300 Full fusion outputs had valid evidence paths, including 91 of 94 Rule 4-only outputs.A correct path required same-investor, same-period evidence, correct product alignment, and an eligible cash-flow signal.Source: §4.5 · p. 9, Table 7 · p. 9

5 Discussion

FIRME’s registry reattachments extend correct matches into cumulative histories rather than duplicate edges. Its lifecycle analysis distinguishes observed evidence trajectories from latent investor preference and examines how parameter choices shape those trajectories.

5 Discussion

  • Correct registry reattachments extend one cumulative evidence history rather than create duplicate Emerging edges; CMR should be read alongside MMR and Merge Precision.An incorrect registry match can route later occurrences to the wrong anchor and history.Source: §5 · p. 9

5.1 Lifecycle Behavior Over Time

  • Observed Market Volatility activity rises and subsides; consecutive unmatched periods shift memory from Emerging toward Stable and Fading, with no Reactivated share in the observed window.These states reflect observed evidence, not latent investor preference.Source: §5.1 · p. 9

5.2 Analytical Parameter Behavior

  • The analytical sensitivity varies saturation speed with α and decay rate with γ; decay begins at the first miss, while N separately sets the Fading threshold.Under defaults, an edge starting at 0.5 saturates after three consecutive matches and retains 51% of its pre-decay score after three misses; periods are monthly for investor calls and reporting-period-based for S&P 500.Source: §5.2 · p. 9, Table 8 · p. 9

6 Conclusion and Future Work

FIRME combines registry-anchored identities with within-period evidence fusion and deterministic lifecycle updates. Its evaluations show improved reattachment against the no-alias control, while future work calls for broader settings and signals.

6 Conclusion and Future Work

  • FIRME increased CMR over the no-alias control by 7.8–12.1 percentage points and achieved 2.3–3.7 times LLM-Norm’s CMR under the judge and fallback-review protocol.The shared human audit found CMR 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.Source: §6 · p. 9
  • Figure 4 covers the Market Volatility topic group from November 2024 through August 2025, showing monthly observed investors and lifecycle-state shares over accumulated memory edges..Source: Figure 4 · p. 10
  • Future work should learn entity- and relation-specific decay, incorporate negative signals, support firm-level aggregation, and test longer windows, more backends, and independent judges and generators..Source: §6 · p. 10

7 Limitations

The paper’s limitations concern component attribution, data and time-window coverage, audit scope, use-case-specific update choices, and model and evaluator generalization.

7 Limitations

  • The alias-memory ablation tests persistent aliases as a combined mechanism and does not separately quantify canonical registry state versus semantic fallback.Lifecycle updates are deterministic given reconciliation decisions, but extraction and semantic matching remain model-dependent.Source: §7 · p. 10
  • The proprietary investor-call corpus and internal product-annotation resource are unavailable publicly, limiting independent reproduction of those experiments.The S&P 500 corpus lacks paired structured cash-flow records, so it tests extraction and temporal reconciliation rather than full heterogeneous fusion.Source: §7 · p. 10
  • Neither evaluation window spans an annual cycle, limiting longer-term seasonality and lifecycle conclusions..Source: §7 · p. 10
  • Human-audit reattachment recall is estimated for sampled incoming units against an adjudicated historical reference, not a complete population reference.It is conditional on extracted observations and does not measure extraction recall or relationships omitted from the original calls; estimates also depend on sampling and reference quality.Source: §7 · p. 10
  • The product–topic ontology, $100K threshold, and lifecycle parameters are use-case specific, and uniform reinforcement ignores relation type and evidence intensity.An unmatched edge may reflect reduced engagement, missing inputs, or no observation opportunity, which the model cannot distinguish.Source: §7 · p. 10
  • Two backends limit model-level generalization, and relying on one LLM judge may introduce evaluator-specific bias.Manual fallback review and human audit may retain annotator subjectivity; shared Opus-generated documents control input variation, though generator–extractor affinity may remain.Source: §7 · p. 10

A Human Audit of Relationship Reattachment

The human audit uses shared, sampled incoming units and an adjudicated historical reference to compare reattachment strategies. Its results and protocol are distinct from the registry procedure, which reconciles entities against the prior-period registry and retains accepted aliases.

A Human Audit of Relationship Reattachment

  • The audit sampled 200 final-period incoming units per dataset–backend setting, including unmatched units, and all five strategies shared the sample and reference within each setting.Units were not deduplicated, and multiple observations could originate from one call.Source: A Human Audit of Relationship… · p. 12
  • Two annotators independently reviewed all 800 units, were blinded to strategy predictions and automated judgments, and searched historical evidence beyond strategy-proposed candidates.Initial substantive agreement was 704/800 (88.0%); 22 of 96 cases requiring resolution received third-person review, and all final labels were resolved.Source: A Human Audit of Relationship… · p. 12
  • MATCH required an evidence-supported historical relationship with the same head and typed target identity; NEW required a history search with no match, while insufficient evidence was UNRESOLVED.Final adjudicated labels supplied the reference for strategy performance.Source: A Human Audit of Relationship… · p. 12
  • Human-audit CMR divides correct reattachments by all 200 sampled units, precision by accepted reattachments, and recall by gold recurring units.Recall is conditional on extracted streams and eligible history; extraction recall over original transcripts is outside the audit’s scope.Source: A Human Audit of Relationship… · p. 13
  • Table A2 reports CMR, precision, and recall with two-sided Wilson 95% confidence intervals; F1 is a point estimate, and unresolved final labels were not excluded.RAW and spaCy-Lemma Table 3 CMR uses population acceptance rates, whereas Table A2 computes CMR directly from the sample, so estimates can differ.Source: A Human Audit of Relationship… · p. 13, Table A2 · p. 13
  • Full FIRME achieved the highest observed human-audit F1 in every setting (76.4–92.7%), exceeding FIRME w/o Alias Memory by 17.2–20.5 percentage points.Table A2 reports performance on 200 shared incoming units per setting; F1 is a point estimate.Source: A Human Audit of Relationship… · p. 13, Table A2 · p. 13

B Reconciliation Strategies and Alias-Memory Control

  • FIRME w/o Alias Memory retains historical canonical identities and semantic fallback but does not accumulate alternative names; full FIRME adds persistent alias reuse.Both FIRME variants retain semantic fallback and evolving registries; LLM-Norm can match equal normalized names or identifiers but does not use historical canonical identities for incoming names.Source: B Reconciliation Strategies and… · p. 14, Table B1 · p. 14

C Cross-Period Entity Reconciliation Pseudocode

  • For periods after initialization, registry reconciliation checks deterministic matches before semantic matches, then adds accepted surface forms to the selected anchor’s alias set.Matching queries read the prior registry; matched targets adopt canonical IDs for edge reconciliation while head IDs remain unchanged.Source: C Cross-Period Entity Reconciliation… · p. 15

D Exploratory Semantic Threshold Selection

The appendix evaluates semantic matching thresholds and specifies deterministic short-term fusion. It selects 0.8 as a tradeoff between precision and match coverage, then describes how corroborated cash-flow signals update the fused graph.

D Exploratory Semantic Threshold Selection

  • The threshold analysis evaluated 5,610 unresolved entities at 0.7, 0.8, and 0.9, with two annotators reviewing 100 accepted matches per threshold.Semantic matching considered same-type registry candidates; the rule-based matcher accepted the highest-scoring candidate meeting the threshold.Source: D Exploratory Semantic Threshold… · p. 15
  • 0.8 was selected to improve precision over 0.7 while retaining more matches than 0.9.Table D1 reports the exploratory threshold analysis; the supplied text does not give its cell values.Source: D Exploratory Semantic Threshold… · p. 15, Table D1 · p. 15

E.1 Rule-Based Short-Term Fusion

  • Short-term fusion retains and groups observations, then labels relationships as strong-interest when Rule 3 or 4 applies, while referencing supporting observations and ontology links.Original observations remain intact, and topic relationships or uncorroborated product relationships retain their original labels.Source: §E.1 · p. 16

E.2 Topic and Product Histories Across Periods

FIRME’s appendices illustrate how relationship histories are updated across periods, how extraction preserves raw evidence alongside canonical fields, and how entity matching and a fixed product–topic ontology support graph construction and fusion.

E.2 Topic and Product Histories Across Periods

  • Each relationship history updates once per period: discussions persist as observations, while an unobserved product relationship decays without historical support triggering another update.A discussion supporting multiple allocated products can be referenced by each product; an ontology link alone does not create a topic observation or reinforce it.Source: §E.2 · p. 16, Table E1 · p. 16, Table E2 · p. 16

F.1 Relationship Extraction

  • Extraction returns linked raw and canonical layers, preserving transcript wording and evidence while recording normalized entities and graph-ready relationship fields.The rules require transcript-only extraction and prohibit correcting grammar, expanding abbreviations, or rewording raw fields.Source: §F.1 · p. 17

F.2 Semantic Entity Matching

  • Semantic matching scores unresolved incoming entities against supplied same-type registry candidates using canonical names and aliases, with thresholding and ID assignment performed programmatically.For FIRME without alias memory, candidate descriptions omit aliases.Source: §F.2 · p. 18

G Product–Topic Ontology: Representation and Use

  • The fixed product–topic ontology contains topical associations, not entity equivalences, and Rule 4 uses those links to connect topic discussion with product cash-flow evidence.The ontology derives from a manually annotated proprietary resource covering approximately 17,000 products and is not publicly available; an ontology link alone does not trigger an upgrade.Source: G Product–Topic Ontology:… · p. 19

H Additional Evaluation Details

The additional evaluation details describe human review of historical reattachments, checks on GraphRAG evidence, and instance-level node and edge scoring.

H Additional Evaluation Details

  • Accepted reattachments were judged against historical relationships and supporting evidence through T−1, with human review replacing incorrect initial judgments after disagreement resolution.Correctness required the same head and typed target identity.Source: H Additional Evaluation Details · p. 20
  • The GraphRAG adapter retained raw mentions, normalized targets, and candidate evidence excerpts, which were checked against source passages located by text-unit identifiers.Normalization was assessed on corresponding raw-to-normalized target pairs.Source: H Additional Evaluation Details · p. 20
  • Node and edge metrics were calculated per sampled instance and macro-averaged; grounded extra recovered items raised adjusted precision but did not count as recovered reference items.Edge matches required aligned endpoints, relation type, and direction to match the reference.Source: H Additional Evaluation Details · p. 20

H.1 Short-Term Fusion Annotation

The short-term fusion audit judged whether relationship outputs had correctly aligned entities, periods, and evidence under the applicable fusion rule.

H.1 Short-Term Fusion Annotation

  • Reviewers checked investor, product, and period alignment, source support, and whether each relationship met its applicable fusion rule.Direct outputs required a Rule 3 path; Full outputs required at least one valid Rule 3 or Rule 4 path.Source: §H.1 · p. 20
  • A disputed VTEB relationship was resolved as correct because Rule 4 permits corroboration through a topic–product association without a direct product mention.The annotators resolved the disagreement by revisiting the evidence and Rule 4 criteria.Source: §H.1 · p. 20
Loading 2610.40482v1…