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

Yunbai Zhang; Yunyi Gao; Han-Tai Shiao

cs.AIOct 29, 2026

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

Authors: Yunbai Zhang, Yunyi Gao, Han-Tai Shiao

Organizations: The Vanguard, Malvern, Pennsylvania, USA

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.

Figures & tables

Panel of Figure 1 — Figure 1. Coupled short- and long-term fusion in FIRME across consecutive periods. Each short-term graph $S^{t}=G_{u}^{t}\oplus G_{s}^{t}$ fuses unstructured and structured graphs. Dashed arrows show registry update from $R^{t-1}$ to $R^{t}$ , conditioned on $S^{t}$ ; solid arrows to $M^{t}$ show memory reconciliation of $S^{t}$ , $R^{t}$ , and $M^{t-1}$ . In FIRME, across consecutive periods, the short-term graph and the previous registry jointly update the current registry (registry update); the short-term graph, the updated registry, and the previous memory graph are then reconciled to form the current memory graph (memory reconciliation).

Entity types
person/org	Tracked investor or asset-management firm
product	Financial instrument (e.g., VOO , BND )
topic	Investment theme (e.g., ESG Investing)
Relation types
discussed [-1pt] Unstructured	Product/topic identified as substantively discussed
cf-signal [-1pt] Structured	Investor shows significant inflow into a product

Table 1. Entity and relation schema for investor-call fusion.

Rule	Type	Antecedent	Consequent
1	Pass	discussed(Alice, VFIAX, 1)	$\Rightarrow$ discussed(Alice, VFIAX, 1)
2	Pass	cf-signal(Bob, VOO, 2)	$\Rightarrow$ cf-signal(Bob, VOO, 2)
3	Fuse	discussed(Chris, BND, 1) $\oplus$ cf-signal(Chris, BND, 2)	$\Rightarrow$ strong-interest(Chris, BND, 3)
4	Fuse	discussed(Emma, TaxEff, 1) $\oplus$ cf-signal(Emma, VTEB, 2) , with $(\texttt{VTEB},\texttt{TaxEff})\in O$	$\Rightarrow$ discussed(Emma, TaxEff, 1) and strong-interest(Emma, VTEB, 3)

Table 2. Short-term fusion rules in FIRME. Rules 1–2 retain uncorroborated relationships; Rules 3–4 assign strong-interest to products using current-period evidence from the same investor. Rule 4 retains the discussed-topic relationship. Original observations and provenance are preserved under all rules; TaxEff denotes Tax Efficiency.

Panel of Figure 2 — Figure 2. Memory-edge lifecycle with inactivity threshold $N=3$ . The entry arrow initializes a relationship as Emerging. 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. A five-state lifecycle diagram. New relationships enter Emerging. A match leads to Reinforced unless the prior state is Fading, in which case it leads to Reactivated. Consecutive misses lead to Stable through the third miss and to Fading on the fourth. Initialization and matches reset the inactivity count. Solid transitions denote matches; dashed transitions denote misses.

Panel of Figure 3 — Figure 3. Topic consolidation under LLM-Norm (a) and FIRME (b); colors link surface variants assigned to one canonical topic, while uncolored topics remain distinct. Two investor-centered topic graphs compare 11 surface-form topics under LLM-Norm with five canonical topics under FIRME. Colors indicate which surface variants are consolidated, while uncolored topics remain distinct.

Dataset	Backend	Strategy	Edge CR	Entity CR	MMR	Merge precision	CMR	Reattachment Recall
Investor calls	Opus 4.6	RAW	0.0%	0.0%	7.6%	100.0% †	7.6% †	15.0%
		spaCy-Lemma	3.0%	7.6%	8.4%	94.1% †	7.9% †	20.0%
		LLM-Norm	6.1%	22.6%	8.3%	98.5%	8.2%	20.0%
		FIRME w/o Alias Memory	15.4%	43.7%	24.1%	92.3%	22.2%	52.5%
		Full FIRME	20.2%	51.1%	31.8%	95.5%	30.4%	77.5%
Investor calls	GPT-5.4- mini	RAW	0.0%	0.0%	5.9%	100.0% †	5.9% †	15.0%

Table 3. Memory usability. CR (compression rates) and MMR summarize population outputs. Precision uses human-audit estimates for RAW/spaCy-Lemma (

\dagger

); otherwise, Gemini 3.1 Pro judges all accepted reattachments, with human review of initially incorrect cases. CMR is population MMR

\times

precision. All reattachment recall values come from the shared human audit (200 incoming units per setting; CIs in Appendix A ). Bold marks the highest CR, MMR, CMR, and recall within each setting.

Strategy	Entity canon.	Semantic matching	Total
LLM-Norm	5.42M	—	5.42M
FIRME w/o Alias Memory	5.29M	1.39M	6.68M
Full FIRME	5.37M	0.83M	6.20M

Table 4. Token consumption for approximately 1,300 investor calls in the final period

T

, using Opus 4.6. Totals cover entity canonicalization and cross-period semantic matching; earlier registry-construction periods are excluded. M denotes million tokens.

Dataset	Extractor	LLM Backend	Hall.	N.Err.
Investor calls	LLM2Graph	Opus 4.6	3.3%	3.2%
Investor calls	LLM2Graph	GPT-5.4-mini	6.6%	7.74%
S&P 500	LLM2Graph	Opus 4.6	7.4%	2.8%
	LLM2Graph	GPT-5.4-mini	5.87%	22.16%
	GraphRAG	GPT-5.4-mini	11.5%	17.0%

Table 5. Extraction-faithfulness Hallucination (Hall.) and Normalization Error (N.Err.) rates by extractor and backend.

Dataset	LLM Backend	Level	Adj. P	R	F1
Investor calls	Opus 4.6	Node	0.952	0.918	0.931
	Opus 4.6	Edge	0.954	0.917	0.929
	GPT-5.4-mini	Node	0.963	0.902	0.926
	GPT-5.4-mini	Edge	0.968	0.898	0.923
S&P 500	Opus 4.6	Node	0.982	0.786	0.865
	Opus 4.6	Edge	0.982	0.783	0.862

Table 6. Round-trip recovery: adjusted precision (Adj. P), recall (R), and F1, independently macro-averaged across instances at node and edge levels.

Condition / output subset	Audited cases	Human-audit fusion precision
Direct fusion (Rule 3)	200	99.0% (198/200)
Full fusion (Rules 3+4)	300	97.7% (293/300)
Added by Rule 4 beyond Rule 3	94	96.8% (91/94)

Table 7. Human audit of short-term fusion on final-period investor-call data using Opus 4.6. Precision uses final judgments after disagreement resolution. The final row is a subset of the 300 Full outputs and is included in the Full result.

Panel of Figure 4 — Figure 4. Market Volatility topic group in the investor-call corpus, November 2024–August 2025: monthly observed investors (left) and lifecycle-state shares over accumulated memory edges (right). Two-panel figure for the Market Volatility topic group on the investor-call corpus from November 2024 to August 2025. Panel (a) is a line chart of monthly observed active investors; panel (b) is a stacked bar chart showing the shares of lifecycle states (Emerging, Reinforced, Reactivated, Stable, Fading) over accumulated memory edges.

Parameter / behavior	Low	Default	High
Reinforcement $\alpha$	0.1	0.2	0.3
Matched periods to reach $1.0$	5	3	2
Retention $\gamma$	0.7	0.8	0.9
% retained after 3 periods	34	51	73

Table 8. Analytical effects of

\alpha

and

\gamma

across low, default, and high settings; bold marks defaults.

Knowledge Graphs

Knowledge graphs represent entities and relationships as structured, queryable networks for integrating data and supporting reasoning.

See all 235 papers

Absorbing Complexity: An Interaction-Native Knowledge Harness for Financial LLM Agents
Jun 1, 2026Ailiya Borjigin, Igor Stadnyk, Ben Bilski +4LLM Agent Memory LLM Agents
True Trading
Point-in-Time Financial RAG with Frozen LLMs and Market-Feedback Adaptive Retrieval
May 29, 2026Zijie Zhao, Roy E. WelschRetrieval-Augmented Generation Quantitative Finance
Massachusetts Institute of Technology · Cambridge, MA, USA
OntoKG-EQ: A provenance-grounded, competency-question-governed knowledge graph for auditable analyst querying
Sep 8, 2026Furqan Nasir, Muhammad Atif Saeed, Muhammad Ehsan +2KG Construction Data Provenance
City University of Science and Information Technology (CUSIT), Peshawar, Pakistan · National University of Computer and Emerging Sciences2026 (FAST-NUCES), Islamabad, Pakistan · University of Engineering and Technology (UET), Peshawar, Pakistan

Jun 1, 2026J/K move · Enter open · S save

Absorbing Complexity: An Interaction-Native Knowledge Harness for Financial LLM Agents

Ailiya Borjigin, Igor Stadnyk, Ben Bilski, Maksym Chikita+3

True Trading

LLM Agent Memory LLM Agents LLM Context Management Finance Memory-Augmented Language Model Agents

Open paper AskPDF arXiv

Financial AI agents often fail for a simple reason: they make users carry the complexity. A user must repeatedly restate goals, risk preferences, portfolio context, past judgments, and shifting market assumptions, while the agent answers, retrieves, acts, and forgets. In finance, this is not just inconvenient. In tasks such as market analysis, copy-trading review, and trade preparation, forgotten context and stale memory can create latency, repeated errors, weak auditability, and unsafe decisions. We propose the interaction-native knowledge harness (InKH), an architecture for financial LLM agents that absorbs complexity into the system. InKH converts user, market, portfolio, and tool events into structured operational knowledge. It uses passive knowledge injection to assemble a bounded working context buffer before the main model step, temporal graph memory for low-latency retrieval, a wiki audit surface for human-readable governance, and background extraction with maturity, decay, and write-time invalidation. We evaluate InKH on a reproducible controlled synthetic benchmark with 24 random seeds, 4 rounds, 80 episodes per round, and 6 baselines, producing 46,080 baseline-conditioned evaluations. InKH achieves mean task quality of 0.815 at 900 ms latency. Compared with agent-driven wiki-walk memory, it reduces latency by 82.95 percent, token cost by 82.29 percent, and stale-knowledge usage by 96.58 percent, while improving quality by 0.108 and traceability by 0.461. Compared with a temporal-graph system without invalidation, it improves quality by 0.050 and reduces stale-memory usage by 96.58 percent with comparable serving cost. The results support a design thesis for financial AI: adoption happens when complexity is absorbed by the system rather than transferred to the user. The benchmark validates architecture-level behavior, not live trading performance.

Abstract

Knowledge Graphs

Explore similar work

Absorbing Complexity: An Interaction-Native Knowledge Harness for Financial LLM Agents

Point-in-Time Financial RAG with Frozen LLMs and Market-Feedback Adaptive Retrieval

OntoKG-EQ: A provenance-grounded, competency-question-governed knowledge graph for auditable analyst querying