I Got Tired of Re-Explaining My Codebase to AI — So I Built a Memory Layer

#ai #productivity #devtools #coding

It's 9:47 AM. I'm reopening my IDE to continue yesterday's work on an auth system. I ask Claude to pick up where we left off.

"What authentication approach are you using? JWT or sessions? Which OAuth provider? What's your database?"

We literally discussed this yesterday. For an hour.

This kept happening to me. Every. Single. Session.

Disclosure: I'm the founder of ContextStream — I built this because I couldn't stand paying this tax anymore.

The problem nobody budgets for: AI amnesia

AI coding assistants are incredible inside a single chat. They can reason about architecture, write production code, catch bugs.

But the moment you close the window? Total amnesia.

After ~18 years of shipping products, I've learned to notice invisible productivity taxes. This one was huge:

Re-explaining my stack
Re-listing architectural decisions
Re-attaching the same context files
Re-arguing patterns we already settled

I started tracking it. I was spending 10–15 minutes per session just getting the assistant back up to speed.

Why the obvious "solutions" didn't solve it

I tried all the usual workarounds:

Chat history: noisy, not portable across tools, and I still had to scroll and re-read.

Built-in memory toggles: tied to one product; I bounce between tools depending on the task.

Pasting context every time: it works, but defeats the point of having an assistant.

What I actually needed was a memory layer that:

Captures decisions as I make them
Retrieves the right context automatically
Works across the AI tools I use

What I built: a memory layer behind my AI tools

I spent the last year building ContextStream — a memory layer that sits behind my AI tools via MCP (Model Context Protocol). MCP is a protocol that lets AI clients call "tool servers" to fetch context.

The core insight is simple:

Storage is cheap. Retrieval is hard.

If you dump everything into context, token costs explode and the model gets confused. The only thing that works is delivering the right context at the right time.

So ContextStream captures three kinds of "project memory":

Decisions — "We're using JWT with refresh tokens"
Context — indexed code/docs so the assistant can retrieve what matters
Connections — which decisions affect which modules

A tiny "before → after"

Before:

"JWT or sessions? Which provider? Which database?"

After:

"Last time we chose JWT with refresh tokens. OAuth provider is X. The auth code lives in …. Want me to continue with the refresh rotation + middleware?"

That's the bar I wanted: start where we left off, not at square one.

Setup (the happy path is one command)

npx -y @contextstream/mcp-server setup

That's it — it configures MCP for the tool you're using.

What actually changed for me

The obvious win: no more re-explaining.

The surprising win: consistency.

Before, my assistant would suggest camelCase on Monday and snake_case on Wednesday. Now it remembers "this codebase uses camelCase" and stays consistent.

And when bugs resurface (they always do), it can pull the previous fix back into view: