Why AI coding agents need persistent memory

Every time you open a fresh conversation with an AI coding assistant, it starts from nothing. It doesn't remember what you decided yesterday, the approach you already tried and threw out, or that your "users" list doesn't actually have an email field. So you explain it all again. And it cheerfully suggests the exact thing that didn't work last time.

That gap — not how smart the model is — is what most often makes an AI assistant feel junior. So I built synaptic: a memory that lets a coding assistant carry what it learned in one session into the next.

It's not the brains, it's the amnesia

A sharp model with no memory is like a brilliant contractor who never writes anything down. It can work out whatever's in front of it, but it can't build up knowledge. Every decision, every correction, every hard-won detail about your project lives only inside the current chat. Close the window and it's gone.

The usual workarounds don't really scale: paste the same background in every time, or keep one giant "read this first" document the assistant only half-reads. What you actually want is for the relevant past context to show up on its own — and nothing else to.

What it remembers

synaptic writes down three kinds of things, and pulls the relevant ones back when a new session starts:

• Decisions — what we chose, and why.
• Corrections — the moments I told it it was wrong, so it stops making the same mistake.
• Project facts — the boring truths that always get lost between sessions: which field actually exists, the command to deploy, the way we name things.

The assistant saves and recalls these itself, as it works. You don't manage it.

Getting the right memories back

Recall is the entire game. Surface the wrong memories and you've just added noise. So synaptic looks two ways at once — by the exact words and by meaning — then blends the results, so a search for "how do we handle logins?" still finds the note you actually wrote as "authentication flow." (I wrote a separate, plain-language piece on how that local search works.) It all runs on your own machine, which for a tool reading your private project isn't optional.

What it refuses to do

It's deliberately not a "second brain" that hoards everything — that just rebuilds the same noise problem on your hard drive. It saves on purpose: decisions, corrections, facts. The value was never in remembering everything. It's in remembering the right things, and bringing them back at the right moment.

The takeaway

Most of the effort to "make the AI smarter" goes into fancier prompts and bigger models. A surprising amount of the real improvement comes from something far more boring: letting it remember. If your assistant keeps relearning the same things, you probably don't need a bigger model. You need a notebook.

synaptic is open source — more on the project page.