A letter from an Echo user: how I turned disposable AI work sessions into compounding intelligence

First pain: once a session gets long, the agent gets dumb.

Something we already agreed on suddenly disappears. It starts repeating mistakes I clearly warned it about, and I have to explain the same thing again and again.

In those moments I think: what am I even doing here? This thing is really dumb.

But it is already the best model. And most of the time it is not like this. So why?

I know I should open a new session, but I cannot. A new session means explaining everything again. It means the agent has to reread my codebase from zero. And the result might not even be better.

Second pain: my memory and experience cannot exist across tools.

What happens in Claude Code is hard to move into Codex.

The research I did in ChatGPT and Gemini, and all those good conversations, are also hard to bring naturally into my product and coding agents.

So I am constantly searching, copy-pasting, and digging through different tools.

And I still often cannot find what I need.

The more AI tools I use, the more it feels like I am carrying my own brain between separate rooms. Time and life get burned in the painful act of looking for things.

Sometimes I think: what if the moment I remembered something, I could get it back instantly?

I have seen all the knowledge graph posts. Honestly, I tried that too. I used to be a heavy Obsidian user and a heavy Notion user. I built my own knowledge graph.

But the maintenance cost was too high, and it did not solve my problem.

I always felt there had to be a better way.

The third pain is the one that cost me the most.

Every day I burn through tokens that would cost hundreds or even thousands of dollars at API prices. I also spend hours sitting in front of AI, slowly building up one stretch of deep, execution-ready understanding.

Inside it are decisions we made, lessons from failed attempts, and accurate repo understanding the AI only built after reading the whole codebase back and forth.

Then the code gets pushed, I go eat dinner, I close the session, and tomorrow I open a new one. The details are gone.

Yes, the old session still contains descriptive traces. You can ask a new agent to reconstruct the details slowly.

But that costs too much time and too many tokens.

And many of the subtle things that were so sharp and alive inside the original conversation cannot really be reconstructed.

Code can go into git.

Where should my thinking process go?

I tried asking the agent to write markdown. The markdown became hundreds of lines long, then longer, then messier.

It felt like taking fresh ingredients and turning them into salted fish and cured meat: slow, tiring, and missing the original nutrition and taste.

By the time work ends at night, I am already tired. I know I should organize things, but I really do not have the time or energy. By the time I remember to do it, all the juice is gone.

It does not care how hard you worked yesterday. It does not care how many tokens you spent. It does not care how precious that instant of judgment was.

It just keeps moving forward.

And the intelligence you fought so hard for yesterday gets left behind in yesterday.

I believe maybe you feel this too.

For agent builders, this problem is so painful that it is hard to say out loud. It is too ordinary. So ordinary that people assume this is just reality.

Then I installed Echo MCP.

Strangely, all three problems started disappearing together.

It tells me how dirty my context window is

Echo sits quietly underneath my workflow and gives me one metric: exactly how dirty each workspace context window is, as a percentage.

You may not know how dirty your window is. If you do not, read this piece. After you read it, you will not be able to sit still.

With live data, I can immediately tell when the agent is about to get dumb.

At that moment, I use Echo to open a new session, and it carries over the context I actually need.

It feels absurdly good.

The dumb agent, the forgetful agent, the agent that repeats the same mistake again and again: that whole problem disappears.

It gives my agent decision-level detail back

I no longer worry that the things I figured out will fail to stay, or fail to appear when I need them.

Echo makes my agent extremely smart. It can grasp decision-level details inside the project:

how a decision was made, what context surrounded it, which doc changed, which file moved. It is clear, and it is fast.

The same is true across projects.

When I am designing a feature and suddenly need to recall user data or user traits, I ask Echo in one sentence. The agent does not have to dig through files.

Markdown can never do this.

With the same limited time, I get much more done. My confidence at work goes up, while my context window gets much smaller.

Most of the time it remembers automatically. Only for especially important things do I manually say, "save this."

The two modes work together well. I spend zero minutes maintaining it.

It is like a portable freshness layer: valuable things go in fresh, and when they come back out, they are still warm.

My conversations are no longer thrown away

Now every dollar I spend feels like it turns into reusable future value.

Even the detours eventually find their proper place.

Here is a real example.

Late one night at the end of June, I finished writing a client-side encryption system. I was so tired that I recorded nothing. Later, all I remembered was: "that was complicated."

Last week I asked Echo about it.

A few seconds later, it brought back the full encryption parameters, several design reasons for which storage handled what, and one thing that made me sit up straight:

a critical risk marked "fix next week" that I had completely forgotten.

Without that recall, it would have stayed there until the day it exploded.

I have never used a memory product this good.

It truly feels like my second brain. And I am no longer worried.

How do I use it? Three steps.

First, install the MCP. Let Doctor scan my workspace once, turning the work I have already produced into memory.

Then I can start.

From day one, I am working with the memory of the past several months.

Second, in normal work, use search memory.

When I need a detail or want an association, I ask Echo, and it finds it for me.

Third, when a session gets heavy, click carry over.

It generates a clean context for me to bring into a new conversation and keep going.

One week later

Here are the numbers first:

The archaeology tax of new sessions, which used to average 116.6k input tokens per startup turn, basically disappears.

The 61.4% historical residue at the back half of long sessions no longer follows me overnight.

The same work dropped from 1.51M input tokens to 0.32M.

The full measurement is in this research piece.

But honestly, the change in the numbers still does not fully capture the change in my inner confidence at work.

The most interesting part is this: I used to feel like a $20 subscription ran out almost immediately. Now it feels remarkably durable.

I save a lot of tokens, and I produce more.

Echo MCP keeps the ideas, judgments, failed attempts, and instincts I create with AI every day from dying that same night.

That changes the texture of work.

I no longer feel like I am restarting my life again and again.

I no longer spend my strength re-explaining, re-finding, and re-teaching the agent who I am.

So I started having a strange feeling:

It is not that I became better at enduring.

It is that my work finally started compounding.

It really feels like I gained an extra piece of working life.

The world was not originally like this.

Originally, everything you produced with AI would slowly sink after the session ended. The judgments, the routes, the failed experiments, the subtle details that only made sense in that moment: time would take them away.

Echo changed that world a little.

It lets those things stay.

It keeps yesterday from disappearing.

I do not know what agents will become in the future. But I am already sure of one thing:

I do not want to use agents without Echo memory anymore.

The good news is that Echo can connect to the main AI workflows I use.

You really should try it.