Project Context, Portfolio Context, and Self-Improvement for AI Agents

Project Context gives each Agent Analytics project a small memory: the goal, the activation events, and the glossary for important event names.

That context travels with analytics results, so your agent can explain growth using the product’s real milestones instead of generic funnel labels.

Update: Agent Analytics now also supports Portfolio Context. Project context stays per surface. Portfolio context is account-level memory for cases where your landing page, directory, docs, app, and other owned surfaces work together as one growth system.

We were inspired by what Nous Research is doing with Hermes Agent: a self-improving agent that preserves useful learning instead of starting cold every time. For analytics, the useful memory is smaller and stricter. Save durable product truth. Skip noisy details.

Two projects can both say “activation” and mean completely different things.

For one product, activation might be:

• sign up for a trial
• create the first item

For another product, activation might be:

• sign up
• invite a team member

For a directory or multi-domain lead operation, activation might not be signup at all. It might be a qualified visitor clicking from the directory into the right product, or becoming a lead on another domain.

Same analytics vocabulary. Different product truth.

Project Context keeps that truth close to the data:

• the goal this project is trying to improve
• the events that really count as activation
• the glossary for what important events mean to the user or business

The value is focus. Your AI is not guessing whether signup matters, whether a team invite matters, or whether a directory click-through is the real milestone. It knows what this project counts as progress.

This is especially useful when one person or team runs many projects: products, docs, directories, landing pages, side projects, lead-gen domains, or microsites.

But this is also where one project-level memory stops being enough.

If agentanalytics.sh is the scanner and conversion surface, openorchestrators.org is a discovery surface, and a directory or ecosystem page is another discovery layer, you need two kinds of truth:

• project context for what counts inside each surface
• portfolio context for what those surfaces are collectively trying to do

Without that split, agents either over-generalize one activation definition across unrelated properties or lose the shared growth story across the portfolio.

Without context, the agent has to infer meaning from event names and whatever was said in the current chat. That works until the conversation ends, a new agent picks up the work, or two projects use similar events for different reasons.

This is a small feature, but it is an important agentic primitive: context per project.

Why Portfolio Context Was Added

Project context answers questions like:

• what does activation mean for this app?
• which event actually matters on this docs site?
• does this directory care about lead clicks or signups?

Portfolio context answers different questions:

• which surface is discovery versus conversion?
• what milestones are shared across the system?
• how should the agent interpret growth across many related projects?

In Agent Analytics, Portfolio Context stores compact account-level memory such as:

• shared goals
• surface roles
• shared milestones
• cross-surface glossary terms

That lets an agent understand that one project exists to generate qualified clicks, another exists to convert, and a third exists to support ecosystem discovery. Different surfaces. One growth system.

The Self-Improving Part

The Agent Analytics skill now treats project context as compact memory. After a Product Growth Scanner run, first instrumentation pass, funnel review, retention review, or human correction, the agent can update what it knows:

• this event means trial activation
• this event means team activation
• this directory domain is trying to send qualified leads
• this project should not use the same activation definition as the app

The next time the agent reads analytics, it starts smarter. Not because the prompt is longer, but because the useful product meaning lives beside the data.

Now it can do that at two levels:

• project context for per-surface product truth
• portfolio context for the shared system across surfaces

Save what should survive the current chat:

• activation definitions
• event meanings
• stable goals
• human corrections

Skip what will go stale:

• this week’s metric value
• one temporary spike
• pasted reports
• guesses the human has not confirmed

How a Human Uses It

The human should not operate a low-level endpoint. The human gives product meaning to the agent.

For example:

For the directory site, activation means a qualified visitor clicks through to the product. For the product app, activation means signup and first project created. Keep those definitions separate when you analyze growth.

The skill handles the rest: inspect the event names, keep the glossary short, update the right project, and use that context next time.

If the same person runs multiple related properties, the instruction can also include the portfolio view:

Treat the docs site as education, the directory as discovery, and the product site as conversion. Track shared milestones like qualified click, signup, and first useful activation across all of them.

The practical habit is simple: when you decide what activation means, tell your agent. When many properties work together, also tell it how those surfaces fit together. Then future analytics answers start closer to the truth.