Institutional Memory as Infrastructure

What happens to an investigation after the investigator quits?

In most banks, the answer is: the case file stays, the judgment disappears. The analyst who spent three years learning which Cyrillic transliterations of sanctioned entity names are genuine matches, which shell company structures recur in Southeast Asian trade finance, which document patterns signal layered ownership, that analyst takes all of it home in their head on their last day.

The case management system has records, outcomes, the final SAR narrative. What it doesn't have is why the analyst decided what they decided, which precedents informed the judgment, or how this case connected to the 400 similar cases the analyst had seen before.

Global financial crime compliance costs $206 billion a year. AML analysts leave every one to three years on average, and backfilling takes four months or longer. Every departure resets some fraction of the institution's investigative capacity to zero.

The stateless investigation

A compliance investigation at a bank works the way a restaurant kitchen would work if nobody wrote down the recipes.

A transaction monitoring alert fires. An analyst opens the case. They pull customer data from the KYC system, transaction history from the core banking platform, screening results from the sanctions vendor, and adverse media hits from a third-party provider. They spend four hours, on average, assembling context that another analyst may have already assembled for the same entity six months ago.

The FFIEC's BSA/AML examination manual warns that banks file SARs with "incomplete, incorrect, or disorganized narratives, making further analysis difficult, if not impossible." The quality problem isn't laziness. Analysts investigating 30 alerts a day with a 90%+ false positive rate produce volume, not depth. When every investigation starts from scratch, the economics push toward speed over rigor.

This is what stateless means in practice. Each investigation is an isolated function call with no access to prior state. The same entity can trigger alerts at the same bank six times in two years, and each investigation begins as if the entity has never been seen before.

What a stateful system looks like

Footprint's Trust Fabric, which we introduced in Part 1 of this series, is the clearest example of institutional memory built as infrastructure. It sits underneath their Percy investigation agent, and its job is to make every investigation aware of every prior investigation.

Three pillars hold it up.

Organizational memory. When Percy completes an investigation, the evidence gathered, the reasoning applied, the data sources consulted, and the human decision at the end all become retrievable precedent. A KYC finding surfaces automatically during an EDD investigation on the same entity. A sanctions screening decision is available to the transaction monitoring workflow.

The example Footprint uses: a sanctioned entity's name appears in Cyrillic across two variant transliterations. A human analyst had previously cleared one transliteration as a false positive. Six months later, Percy encounters the second transliteration on a different case. Trust Fabric connects the two, surfaces the prior clearance, and flags the discrepancy. The agent catches what a stateless system would miss because the stateless system has no way to know that the two names are related across separate cases.

Agent assurance. Trust Fabric tracks whether Percy follows policy correctly. QA teams can inspect the exact policy version, the evidence chain, and the reasoning trace behind each finding. Every agent change is versioned and tested against historical cases before deployment.

Cross-signal pattern detection. Individual cases generate isolated signals. Trust Fabric connects signals across customers, cases, and compliance functions. A pattern that no single investigation would surface, like the same beneficial owner appearing across 12 entities flagged in different compliance workflows over 18 months, becomes visible when the memory layer aggregates.

The feature store analogy

Engineers building ML systems hit a version of this problem a decade ago.

A fraud model trains on a feature called "average transaction amount over 30 days." During training, the feature is computed from a warehouse query. In production, the same feature needs to be available in 10 milliseconds. Two implementations of the same feature drift apart, and the model's accuracy silently degrades.

Feature stores solved this by computing each feature once and serving it consistently to both training and inference. Compute once, reuse everywhere, maintain provenance.

Trust Fabric applies the same principle to investigative judgment. An analyst's conclusion about an entity's ownership structure is computed once during an investigation and made available to every future investigation touching that entity. Without the memory layer, two analysts investigating the same entity six months apart might reach different conclusions because they started from different evidence sets. That's investigation skew, the compliance equivalent of training-serving skew.

But the analogy has a limit. Feature store values are mathematical. A 30-day transaction average is deterministic. An analyst's judgment on whether an ownership structure is suspicious is not. The memory layer has to preserve judgment with its full context: what evidence was available, which policy was in effect, who made the decision, and what they decided. Strip any of those and the precedent becomes a bare conclusion without support, which is worse than no precedent at all.

What Westlaw knew first

The legal system has done structured precedent retrieval for centuries.

When a lawyer researches a question, they search Westlaw or LexisNexis for prior cases with similar facts. Each case has a citation, a court, a date, a holding, and a reasoning chain. The system of stare decisis makes prior judgments explicitly reusable under defined conditions.

Compliance has never had this. Legal precedent is published and case law is public. Compliance precedent is institutional, trapped in one case file in one bank's case management system. Maybe in the analyst's notes. Maybe in their head.

Trust Fabric is an attempt to productize what law firms have done informally through culture and mentorship: preserving institutional judgment in a form that new practitioners can retrieve and apply. The difference is that a law firm's institutional memory lives in partner brains and associate training. A compliance platform's institutional memory needs to live in infrastructure, because the analysts turn over too fast for culture to carry it.

The provenance problem

Memory without provenance is a liability.

If a regulator asks "why did you clear this entity?" and the answer is "because the system said a prior investigation cleared them," the next question is: who conducted that prior investigation, under what policy, with what evidence, and is that evidence still current? A bare precedent without its chain of reasoning creates more risk than it resolves.

This is where regulation creates a hard constraint. In April 2026, the Fed, OCC, and FDIC issued SR 26-2, replacing the 15-year-old SR 11-7 model risk framework. The new guidance covers traditional models and non-generative AI.

It explicitly excludes generative AI and agentic AI from scope.

Banks running agentic compliance systems like Percy have no federal model risk framework governing those systems. The agencies said they'll issue separate guidance "in the near future." In the meantime, examiners are asking about AI oversight during routine exams, with no formal standard to examine against.

The EU is further along. The AI Act's Articles 12 and 13 require automatic recording of events during an AI system's lifecycle and the ability for humans to understand the reasoning behind outputs. A log without an explanatory chain doesn't satisfy the requirement.

So the memory layer needs to carry provenance at every level: which analyst or agent produced the finding, what evidence was consulted, which policy version was in effect, what the decision was, and when. Without that chain, an institutional memory system becomes an institutional liability.

What we can't see

Footprint hasn't published Trust Fabric's architecture. The data model, whether it's a graph database, a vector store, or something else, isn't public. The retrieval mechanism, whether precedent lookup uses semantic search, structured queries, or a hybrid, hasn't been disclosed. The underlying model provider is unknown.

This matters because the architecture determines the failure modes. A vector-based retrieval system is good at finding semantically similar cases but can surface false analogies. A structured graph is precise but brittle when entity relationships don't fit the schema. How Trust Fabric handles precedent conflicts, when one prior investigation cleared an entity and another flagged it, is an open question.

What's visible: the system is in production at FDIC and OCC-regulated banks, at fintechs like Bilt, Nuvei, and MoonPay. It holds SOC 2 Type II, PCI DSS, and ISO 27001 certifications. The Cyrillic transliteration example is presented as a production case, not a demo.

The pattern is real even if the internals are opaque. The question for any team building compliance infrastructure is whether institutional memory belongs in the application layer (built into each investigation tool) or the infrastructure layer (a shared service that all compliance functions read from and write to). Footprint is betting on infrastructure. The alternative, memory embedded in individual tools, fragments the precedent base the same way siloed case management systems fragment it today.