Your AI Doesn't Remember Anything. Here's How to Fix That.

Every AI tool I've built for financial services clients hits the same wall eventually: the AI doesn't remember.

You can build a great RAG pipeline. You can connect it to proprietary data. You can get the compliance layer right, the personalization right, the delivery cadence right. But the next time the system runs, it starts from scratch. There's no accumulated understanding. No institutional knowledge that compounds over time.

Your wealth advisor remembers that a client mentioned their daughter's college timeline three meetings ago. Your AI doesn't. Your compliance officer remembers that a specific phrasing was flagged last quarter. Your AI doesn't. Your analyst remembers the pattern they noticed across four different portfolio reviews. Your AI doesn't.

Memory is the missing layer in enterprise AI. And I recently found an open-source project that addresses it head-on.

GBrain: A Knowledge Brain for AI Agents

Garry Tan — the CEO of Y Combinator — just open-sourced a project called GBrain. It's a personal knowledge brain built on Postgres, pgvector, and hybrid search. The core idea: take everything you know — meeting notes, research, client information, strategy documents — chunk it, embed it, and make it searchable by meaning, not just keywords.

What makes it interesting isn't the vector database. It's the knowledge model.

Every page in GBrain follows what Tan calls the "compiled truth + timeline" pattern. Above a separator, you have your current best understanding of a topic — a living document that gets rewritten as new information arrives. Below, you have an append-only timeline of evidence — the facts, observations, and source material that never get edited, only added to.

That distinction matters enormously for financial services. In our world, you need both: the current answer AND the audit trail of how you got there.

Why This Matters for Wealth Management and Asset Management

When I look at GBrain's architecture, I see building blocks that map directly to problems my clients face every day:

Client Intelligence That Compounds. Every wealth management firm has client information scattered across CRM notes, meeting transcripts, email threads, and the advisor's head. GBrain's model is exactly how a good advisor already thinks about a client relationship. "Here's what I know about the Johnsons right now" (compiled truth) backed by "here's every interaction that shaped that understanding" (timeline). An AI agent that maintains this kind of living client profile transforms meeting prep from a 30-minute scramble into a 30-second query.

Research Memory. Investment research is cumulative. An analyst's view on a sector is built from hundreds of data points gathered over months. Today, most of that context lives in the analyst's head or in PDFs nobody searches. A GBrain-style system where the compiled truth on a sector gets updated every time new research is ingested — with citations back to the source material — turns institutional knowledge from a human dependency into organizational infrastructure. When your best analyst leaves, the knowledge doesn't leave with them.

Compliance Knowledge Base. Compliance rules evolve. Interpretations change. What was acceptable phrasing last year may not be this year. A knowledge brain that tracks the current rule (compiled truth) alongside every regulatory update, policy change, and enforcement action that informed it (timeline) gives compliance teams a searchable, version-controlled source of truth. More importantly, your AI content generation systems can query it before producing anything client-facing.

Deal Pipeline Intelligence. For asset managers evaluating deals, the ability to query "what do we know about grid infrastructure investments" and get results spanning meeting notes, research, portfolio data, and previous deal memos — ranked by semantic relevance, not just keyword match — is the difference between two hours of context assembly and a 10-second answer.

The Hybrid Search Advantage

GBrain solves a problem I've hit repeatedly: keyword search misses conceptual matches, and vector search misses exact names and phrases. Ask "what are our biggest risks right now?" and keyword search returns nothing because the word "risks" doesn't appear in your board prep notes. Vector search might find the right documents but miss the one that mentions a specific competitor by name.

GBrain uses Reciprocal Rank Fusion (RRF) to combine both approaches, plus multi-query expansion via Claude to catch phrasings you didn't think of. The result is search that actually works the way you'd expect — find things by what they mean, not just what they say.

Running It With OpenClaw

Here's where it gets practical. GBrain is designed to work as an OpenClaw skill. You install it, point it at your knowledge base, and then interact with your accumulated knowledge through WhatsApp or any messaging platform OpenClaw supports.

Imagine texting your AI assistant: "What do we know about the Henderson family's estate planning situation?" and getting back a synthesized answer drawn from meeting notes, advisor comments, document uploads, and previous briefings — with citations to the source material.

Or: "Give me a briefing for my meetings tomorrow" and getting context on every client you're seeing, pulled from a brain that's been accumulating knowledge about them for months.

That's not a chatbot answering a question. That's an AI with institutional memory.

What I'd Build With This

For my clients in financial services, I'm thinking about GBrain as a layer underneath the AI systems we already build. Insight Studio (our briefing platform) generates personalized client content — but right now, each generation cycle starts fresh. Adding a GBrain-style memory layer means the system gets smarter over time. The briefing it generates in month six is better than the one it generated in month one, because it has six months of accumulated context about each client, each market theme, and each advisor's preferences.

The compiled truth + timeline pattern also maps beautifully to compliance. Every piece of content our system generates could be stored as a timeline entry, and the compiled truth for each compliance rule gets updated as policies evolve. An auditor could trace any generated content back through the full evidence chain.

I'm going to do a deeper technical walkthrough of this in a YouTube video soon — showing how to set up GBrain, customize it for a financial services use case, and connect it to an OpenClaw agent. If you want to see that, follow along.

GBrain repo: github.com/garrytan/gbrain