Frozen Knowledge vs. Living Context (And Why You Need a Second Brain)

You're smart. But try a quick thought experiment: I take away your calendar, your to-do list, and your notes for one week. (I'll give them back, promise.)

You'd still be equally intelligent, but it wouldn’t take long to realize how much we (humans) lean on external tooling to function at an optimal level.

There’s a simple lesson in that.

You stay on top of your work by checking tools that store and manage context for you: your calendar says when things happen and your to-do list says what needs to get done. Your text messages organize the context of your relationships.

Your AI has the same needs in a purer form. It is smart and capable right out of the box, but its knowledge of the world is frozen.

The way you keep your AI sharp is the same way you keep yourself sharp: you connect it to context, and to tools that manage context for you.

Speaking of which, I’m working on a new product called YouSpot (it’s part of HubSpot Next). It’s a solo CRM (built for the one person company). It’s key feature is a “second brain” which is knowledge-store that you can connect various data sources to. It’s in private beta now (you can put yourself on the waitlist) or you can skip the line (ssh, don’t tell anyone) and buy it for the ludicrously low introductory price of $1/month (vs. $25/month) -- while supplies last. 🙂 Sorry for the self-promotion, but it's relevant, and it's a product I use every day. More details soon.

Anyways, back to the article. Today, I want to break down:

—@dharmesh

Do you know where AI models get their intelligence from?

Modern AI models are pre-trained: they consume a massive snapshot of content (mostly text) and learn to predict what comes next. That process is where the raw intelligence comes from. [Note: Frontier model companies pick which content they’re going to consume, so there is the risk of some biases being introduced into the model in this pre-training period.]

Then they're post-trained, meaning tuned on examples of what helpful answers look like. That second phase turns a very good text predictor into an assistant you can actually talk to.

After that, the model is frozen. Today’s models have a knowledge cutoff. I've written about AI's memory problem before: every new conversation starts from that same frozen snapshot, no matter what you explained yesterday.

Anything AI knows about you, your work, or how you want things built must be derived from context you provided it. Getting personalized results from AI starts and ends with providing additional context: static context and living context.

Not all context is created equal.

Think about the information that keeps you functional. Some of it is stable reference material, and some of it is state that changes every single day. Your AI needs both kinds, and they work differently.

Frozen knowledge is the stuff that rarely changes. Who you are, who your audience is, how you like your output formatted. If you set this up one time, every future conversation starts smarter, and you only edit it when your situation actually changes.

The setup for static context like this really is a one-time job -- though it requires ongoing maintenance. ChatGPT and Claude both have a spot in their settings for standing personal instructions -- I've written a full newsletter on ChatGPT's Custom Instructions before. I highly recommend taking a quick detour to get that set up if you haven’t already.

(For deeper work, both apps also support Projects, where you can add project-level instructions plus uploaded reference docs: a style guide, a product one-pager, an "about me and my business" doc.)

Living context is state. It answers, “Where do things stand right now?”.

Where static context could say “I run a twelve-person content agency for B2B software companies”, living context might leverage your calendar to see who your agency is pitching today, potentially even assemble a pitch deck by stitching together each person’s individual contributions.

If frozen knowledge is an artifact -- usually a written record of an event or preference -- living context is a connection to a frequently updated source of truth. Connecting an AI model to your Gmail account, for example, provides better context into your inbox than manually reporting the state of your inbox.

Official connectors can link your calendar, email, and files, so the model checks your actual current state instead of a stale memory of it. Web search does the same job for what's happening out in the world.

A brief aside on memory:

OpenAI’s ChatGPT and Anthropic’s Claude both automatically create memories about you as you chat, and it's tempting to file that under living context, since it updates on its own.

But if we look at what memory actually stores: distilled, slow-changing facts, like "prefers bullet points" or "works in marketing.", we see that memory is frozen knowledge the model maintains for you.

You can check (and edit) what memories your AI has automatically saved about you:

🤖 Using Claude

🤖 Using ChatGPT

The most popular plugins on ChatGPT (link) or Claude (link) are a good starting point for getting into living context.

If you can’t tell whether living context is required for a given situation: ask how often does this information change?

If the answer is rarely, it's static, so set it once and forget about it for a few months. If the answer is daily or weekly, it's living, so it’s best to find a way to give your model access to the data.

You need both kinds of context, static and living, to get the most out of AI.

Longtime readers have heard me say this before: context is queen. The models will keep getting smarter, but every one of them will still show up knowing nothing about you until you supply them with the appropriate context.

It’s why I’m so excited about YouSpot. I needed a way to pull all the important information in my business together into one, AI-native place.

I'd love to hear from folks who’ve recently experimented with living context -- including if you check out the Second Brain feature in YouSpot. I’m open to any and all feedback. We're making daily updates, so all feedback gets read (and often shipped).

Hit reply and tell me your thoughts -- reading your responses is my favorite part of writing this newsletter 🙂

—Dharmesh (@dharmesh)