Memory System
- Jeremi Karnell
- Jun 22
- 6 min read
The Cold Start Problem
When we were honest with ourselves — really honest — we had to admit that VastAdvisor's MVP had a cold start problem: every time an advisor opens the platform to build a new campaign, the AI starts from scratch.
For a solo RIA running three campaigns a year, that's friction. For a wealth enterprise running dozens of campaigns across a network of advisors, that's a compounding tax on every workflow the platform touches. Every re-explanation is time the advisor isn't spending on clients. Every re-entered preference is a signal the platform should have retained. Every cold-start recommendation is an AI operating on assumptions that contradict what the firm has already taught it.
We built VastAdvisor to be the growth operating system for wealth management. An operating system that forgets everything the moment you close the window isn't an operating system. It's a very sophisticated notepad.
The platform wasn't amnesiac because we didn't care. It was amnesiac because memory is genuinely hard to build correctly in a regulated industry. You can't just log everything and call it memory. You need memory that knows whose firm it belongs to, what level of the organization it came from, and whether it's still accurate or has drifted. You need memory that a compliance officer can stand behind when a regulator asks what the AI knew at the moment it made a recommendation.
None of that is trivial. But none of it is optional either.
What VastAdvisor's Memory System Actually Is
There are three layers, and it helps to understand them separately.
The first layer is what you experience in Eva today. Eva is VastAdvisor's AI copilot — the conversational layer that runs inside every part of the platform. When advisors started asking for memory-like behavior, we didn't wait for the full substrate to be built. We shipped a working memory interface directly into Eva. Type / in the Eva composer and a command palette opens: /recall surfaces what Eva remembers about your firm and current context; /palace opens your Memory Palace — a curated collection of firm-specific facts, preferences, and decisions you've explicitly saved; /pin adds something to that palace; /forget removes it. Type @ and every campaign, lead, ICP, audience, and contact in the platform surfaces as a linked chip you can drop directly into your prompt without re-explaining context you've already established.
The second layer is what VastAdvisor IQ captures as it watches your platform activity. This is what you're looking at in the knowledge graph. Every campaign, audience, lead, pipeline event, rejection, AI execution, piece of market intelligence, and notification is a node. The relationships between them — which ICP drove which campaign, which theme was rejected and why, which audience produced which pipeline signal — are the edges. That is the platform's memory of an entire firm, rendered as a living structure.
The third layer is the Tenant Memory Foundation — the architectural substrate that will make the first two layers fully persistent, bitemporal, and authority-aware across the entire three-tier hierarchy. This is what turns the Memory Palace from a user-managed feature into an automatically accumulating firm intelligence layer. That work is underway, and I'll get to where it's headed.
The Knowledge Graph — What You're Looking At
The VastAdvisor IQ view has a filter rail across the top: All wings, Enterprise, Firm, Audience, Campaigns, Pipeline, Rejections, AI Executions, Knowledge, Market Intelligence, Notifications.
Each of those wings is a domain of platform memory. When you click "Campaigns," the graph highlights every campaign node and the threads connecting them back to the ICPs that generated them, the audiences they targeted, the compliance checks they passed through, and the pipeline signals they produced. Click "Rejections" and you see every theme, headline, or creative the AI generated that an advisor pushed back on — along with what replaced it and what the downstream performance was. Click "AI Executions" and you're looking at a map of every moment the platform's AI touched a workflow, what it knew at that moment, and what it produced.
This is model health and fine-tuning at the firm level. It's not a report. It's not a dashboard widget. It's the AI showing you its own reasoning history — in a form that's navigable, filterable, and zoomable. The 2D and 3D views let you orient to the full structure; the Force (loose) layout spreads the graph to reveal clustering patterns you wouldn't catch in a table.
When an enterprise operator opens VastAdvisor IQ with "Firms: All" selected, they're seeing the combined memory of every firm in their network, in a single structure. Which firms have rich campaign histories with dense audience connections. Which firms have thin graphs — sparse nodes, few edges — and therefore need attention. Which rejection clusters are appearing across multiple firm nodes, suggesting a messaging pattern worth addressing at the enterprise level.
No analyst had to produce that. The platform produced it from its own accumulated activity.
How Eva Carries Your Context
Before the full substrate ships, Eva is doing real work as a practical memory layer.
The Memory Palace is exactly what it sounds like: a named, navigable space where advisors park the facts, preferences, and decisions that should follow every conversation. "Our firm voice is direct and confident, never alarmist." "Primary ICP is the pre-retiree business owner, 55-65, $2M–$10M investable assets." "We do not run content that mentions competitor firms by name." These aren't just notes. When you /recall in a campaign or compliance context, Eva loads the palace before responding — so its recommendations start from what you've already established, not from a blank prior.
The @ mention system closes the context gap at the entity level. When an advisor is asking Eva about a specific campaign, they type @Q2-Retirement-Awareness and the campaign record — theme, targeting, performance history, compliance status — becomes part of the prompt without the advisor having to describe it. Same for ICPs, leads, audiences, contacts. The entity is the context.
And as of v1.11, clicking an @ chip in Eva's response navigates directly to the relevant page. The AI doesn't just remember; it walks you there.
What We've Shipped
Memory has been part of VastAdvisor's development in nearly every sprint since launch. Here's the specific build sequence.
v1.06 — Eva v2 Foundations (Apr 2026). Eva moved from a generic RAG model to workspace-scoped filesystem skills with page-context awareness. Tenant-scoped skills meant Standard and Enterprise tenants started seeing different Eva knowledge bases. This was the architectural pivot — from Eva as a generic chatbot to Eva as a firm-aware guide.
v1.07 — AI Governance Console (Apr 2026). The full AI Governance Console shipped: workflow registry, run history, experiment matrix, kill switch, and PII boundary enforcement. The platform now had a first-class surface for understanding its own AI execution history — the precursor to the knowledge graph's AI Executions wing.
v1.09 — Notification-Driven Eva (May 2026). "Ask Eva About This" launched: any platform notification can now trigger a pre-seeded Eva conversation loaded with the full context of whatever alert fired. Eva went from a tool you had to proactively open to a guide that meets you at the moment something matters.
v1.11 — Memory Palace & Slash Commands (May 2026). The Memory Palace shipped. Seven slash commands live in Eva: /recall, /facet, /go, /palace, /pin, /forget, /help. Eva @ mentions let advisors drop entity-linked chips directly into prompts. ICPs became first-class objects inside Eva — a direct line between a client segment and the AI making recommendations about it. Click-through navigation from Eva chips to platform pages closed the loop.
v1.12 — Content Intelligence Foundation (May 2026). The Campaign Builder overhaul brought the Marketing Strategy document and GTM Playbook into a persistent campaign workspace — a key step toward making campaign intelligence durable, not session-bound.
v1.13 — Theme Intelligence & Regulatory Audit Trail (June 2026). ICP-aware signal filtering — incoming prospect signals ranked against each advisor's ICP before they surface — made the memory-to-action loop bidirectional. And the 7-year regulatory audit trail, with full provenance logging for every CRM contribution and AI write-back, answered the compliance question that wealth firms have been unable to answer until now.
Comments