Implementing Gbrain for a company — Part 1

Why Gbrain (Pros and Cons)

Information in a company is scattered across slack/teams, emails, call notes, SaaS tools and shared drives. From 2022–25 we’ve seen several AI tools that promise to put this alltogether in a single place and make it accessible via AI, companies like grw.ai, gong, clari etc.

Their approach is similar, they connect to all your work tools and ontop of this they have their proprietary LLM workflows. It uses the latent intelligence sitting in your workspace tools and collates them to answer questions or run automated workflows like prediction or call analysis etc.

We find that this approach falls short. Managers often find that the answers are superficial and dont go deep enough. Also these tools are a one size fits all, while a companies workflows are the outcome of years of evolvution influenced by various tools and people along the way.

Gbrain allows you to build custom implementations of an otherwise complex concept called Knowlede Graphs.

Keyword search helps you find exactly what you typed in.

Vector search helps you find what you meant to find.

Knowledge graphs help you find things you might not have known or thought about.

Creating and maintaining a knowledge graph is very complex. Gbrain leverages the main agent of AI harnesses like Openclaw, Hermes or Claude Cowork to abstract away the complexities of setting up and maintaining a Knowledge Graph.

Setting up and maintaining Gbrain though simpler than a traditional knowledge graph, still takes more time than you would spend with an off the shelf AI tool. However, it will be worth it in the long run as it becomes an internal moat. You invest in the likelihood of adoption and a multi-fold jump in productivity.

Gbrain’s best quality can sometimes be it’s shortcoming too. Its just too automated. It can create pages and backlinks for things that you may not actually think are useful. Depedning solely on the main agent, may result in your brain becoming too noisy and degrading the quality of retrieval. I prefer narrowing scopes of what i want gbrain for, and creating my own hybrid (deterministic + main agent evaluation) workflows to keep gbrain updated.

Step 0: Install Gbrain in your Openclaw

Why we chose Openclaw over Claude Cowork or Hermes is another article. But for now, yes, we chose to host a virtual machine on a server with openclaw on it. In the root directory I did the CLI install for Gbrain

Step 1: Recipies

Recipies are setup prompts that tell the main agent how to setup ingestion paths. email to brain, calendar to brain etc. It also covers the backfilling. i.e. consuming and pushing all the older things into Gbrain.

You dont have out of the box recipies for things like salesforce, teams, slack or hubspot, so I had to make my own custom implementations of these, which Ill share once my implementation matures. However, here are a few things I’ve learnt.

• Have a point of view

  : Map out how you want your gbrain to look. I started out building the brain for Sales. So deals were our anchor point. Notes on deals is what most people would search for, all other notes for companies, people, meeting notes, etc should be mapped back to deals.

• Keep it clean

  : We put a stake in the ground and chose a point of actual interest i.e last 6 months. Sales cycles dictates that we needed atleast 6 months worth of conversations and anything before that isnt really relevant.

• Chose a single source of truth:

   Gbrain is configured to create notes on people , companies, deals etc if it the main agent thinks its “notable” and needed. Id avoid this. We chose hubspot as that deterministic place. Only companies, deals and People created in hubspot are created in Gbrain. connected to them we setup deterministic workflows to enrich them with email, slack and

I started off with Hubspot to Brain, Email to Brain, Slack to Brain , and meeting notes to brain.

For recipies, I started off with using existing recipies and built my own taking those as reference. Once you’ve created and saved your recipies in a folder on the VM. You ask you openclaw main agent (who I connected with telegram to talk to) to follow those recipies.

You openclaw will then follow those recipies and create gbrain notes for you. This entails 3 processes.

1. Collection

   : Collect the required notes and materials from the various sources

2. Enrichment

   : Using the instructions provided in the recipies, the main agent creates notes, adds details in the note, and baclinks it into other notes. This is the first version of your knowledge graph.

3. Promotion

   : Gbrain uses a hybrid of BM25 key word search, vector based embedded search and Knowledge Graph traversal, to find information for you. To do the first 2, Gbrain uses a PGSQLlite DB into which it stores the embedding. This is an important step in the recipie, so dont forget it.

Step 3: Scheduling

The recipies are the setup. It creates the base, and ensures that all previous information is backfilled into your Gbrain. Now we need to keep it fresh. Scheduling should also be a part of the recipie. However, I’ve found that the main agent sometimes skips that step. The recipie is essentially a very big prompt telling the main agent what to do. So it’s best that you manually check with the main agent that the scheduling is in place. Also check the end to end automation flow for the scheduled job. The main agent may often skip steps between collection → enrichment → promotion. Also ensure that there’s a step to commit to the brain’s git. Gbrain does create a separate git branch for the brain, which it uses for version control. If there are changes, that the scheduled job did not commit to git, your jobs may fail.

Step 3: Iteration

Once you’ve gotten the gbrain setup and running on a regular basis, then comes the somewhat painful process of iteration.

Ive found plenty of issues with how the main agent chose targets, how it summarized, or created timelines, types of metadata it stored etc. This is a painful process of chosing a few candidates and checking if the correct information has been added and in the intended design.

Some tricks that have proven useful:

• Remote desktop

  : Setup a remote desktop so it becomes easier for you to visually navigate and inspect the notes. This for me is a big productivity boost

• Golden Set

  : Chose a handful of notes that you are confident on knowing what you want.

• Ingest one source at a time:

   I started with Hubspot, then email, then moved on to Slack and Notes. Do it one at a time.

• Instructions should be a prompt:

   I ask that the instructions to the main agent on how to enrich the note, is best when it sits outside the script file. Makes it easier to see how the main agent is reviewing the file and how its making the note.