Your AI isn't Christopher Columbus. Give it a map

We rebuilt the knowledge system underneath our agent, its ability to discover relevant context moved from 84 to 97 across multiple golden test sets, and the same internal comparison ran 27 times faster.

We ran identical code and questions on the same machine, the difference was the context the agent received - how the system organized what we knew and found the relevant pieces before the model started working.

An agent's ability is governed by its context, improve the context - and every task downstream gets better.

Before the rebuild, relationships between facts could remain scattered across documents, so finding the right file didn't necessarily give the system the connection it needed.

We gave those facts a stable place, then mapped what connected to what.

Under the hood we used Postgres, pgvector, and a lexical ontology, the technical setup matters less than the job it was doing - helping the agent discover relevant context.

We haven't solved client-data isolation yet - that part is still planned.

You can test this with one question tied to a decision you actually make.

For example: "Which accounts should sales call first, and why?"

Before you ask it, write down:

• Required facts - what a useful answer needs to know
• Expected reasoning - why the recommendation should make sense
• Allowed sources - where the agent can look
• Data boundary - which customer records it can access

Ask the question five times and save every answer.

You're not looking for identical wording, you're checking whether the agent finds the relevant facts and uses them to reach a consistent recommendation.

Then change how it retrieves context, and run the same question five more times.

In GTM work, this means an agent can retrieve the right account history, campaign context, and decision criteria before it drafts.

You still review the judgment calls and spot-check the research, but better retrieval means less time rebuilding context for routine tasks.

If you need a place to start, AgentMemory is an open-source project built around persistent memory, hybrid retrieval, and knowledge graphs.

It didn't power our internal test - use the repo to inspect how it handles memory and retrieval before you build your own.

Follow the project's installation guide, add one golden question, then compare the answers before and after your change.

Where I can help you:

• YouTube - see the GTM systems and experiments we're running
• Legion - join the waitlist for the GTM operating system we're launching soon
• LeadGrow - apply for a free campaign and let us prove the funnel with you

- Mitch