Firm Memory and private RAG for financial services.
Without Firm Memory, every AI answer starts from zero. With it, the OS compounds. This is the layer that turns a subscription into an asset.
The retrieval and knowledge layer of an AI OS that stores a firm's own documents, deals, positions, chats and decisions with provenance and permissions, so that any agent, workflow or skill can ground its answers in the firm's own signal.
The moment a firm asks 'why didn't the AI know that?' — that we cover this sector, that this client was flagged three months ago, that this filing is our house benchmark — is the moment it discovers it has no Firm Memory. Frontier models know the internet. They do not know your firm. Firm Memory is the layer that fixes that.
What Firm Memory contains
- Documents: research notes, deal memos, credit committee minutes, KYC files, complaints, contracts, filings the firm cares about.
- Structured signal: positions, trades, deposits, withdrawals, CRM interactions, MI, compliance logs.
- Decisions: not just the outcome, but the reasoning trail — 'we passed on this deal because…' — captured with source and author.
- Metadata: provenance (who authored it, when, from which system), permissions (who can see it, at what resolution), freshness (is it stale?).
Why generic RAG fails in finance
The public playbook — 'embed all your PDFs into a vector database, query it, done' — collapses on contact with a regulated firm. It ignores permissions, so the AI happily surfaces a document one desk should never see. It ignores freshness, so it quotes a policy that was superseded last quarter. It ignores structure, so it treats a KYC file the same way it treats a marketing brochure. Firm Memory done properly is a knowledge system, not a vector index.
A firm's data is not a corpus. It is an organism. Firm Memory is the layer that treats it as one.
The shape we ship
- A structured knowledge graph for entities the firm cares about — clients, counterparties, instruments, funds, deals.
- A vector layer indexed off the same entities, so retrieval joins semantic similarity to structural context.
- A permissions model that mirrors the firm's org chart, not a flat access rule.
- A provenance layer that puts a source, author and timestamp on every fact returned.
- A freshness policy that quietly deprecates stale content.
Where to start
Every KJ Capital engagement begins with a two-week £15k AI Diagnostic. We map your existing data estate, model your firm's OS across the six layers above, cost the build, name the first three skills to ship, and hand you a plan your board can act on. It is deliberately priced to be a rounding error against the value it unlocks.
FAQ
How is this different from a vector database?
A vector database is a component. Firm Memory is the layer that combines it with a knowledge graph, permissions, provenance and freshness. Buying a vector DB gives you storage. Building Firm Memory gives you a system.
Do we have to move all our data to build this?
No. Firm Memory reads from source systems in place through connectors and only mirrors what needs to be retrievable. The firm's system of record does not change.
What about regulator questions on data leakage?
Firm Memory is private by construction. Nothing crosses a vendor boundary except embeddings, which can themselves be generated on private infra where the classification requires it.
Start with the £15k AI Diagnostic
Two weeks, one costed OS blueprint, three named skills to ship first.