How memory works.
Neuron doesn't paste your question straight into an LLM. Every request runs through a pipeline that classifies intent, retrieves the memory that matters, assembles a grounded prompt, and — only when you've configured more than one model — routes it to the cheapest model that can do the job.
ClassifyProject MemoryFramework DocsPrompt AssemblyModel Routing
1
Classify
A lightweight classifier reads your question first — not to answer it, but to figure out what it actually is: a quick lookup, a multi-file edit, an architecture question. That classification decides what happens next.
2
Retrieve project memory
Neuron queries your project's Engram memory — working memory and the long-term graph — for facts, decisions, and patterns relevant to this specific question. Not your whole codebase history, just what matters here.
3
Retrieve framework docs
In parallel, Neuron checks its framework documentation memory — auto-fetched docs for the libraries your project actually uses — for anything relevant to the question.
4
Assemble the prompt
Classified intent, retrieved project memory, retrieved framework docs, and your actual question get merged into one grounded prompt — not four separate calls, one structured request.
5
Route to a model
If you've configured more than one model, the assembled prompt's complexity decides the tier: simple asks go to a cheap, fast model; complex ones go to your best model. If you've only set up one model, there's nothing to route — every request goes straight to it.
What the prompt actually looks like.
A real example, assembled by the pipeline above, before it's routed to a model.
[SYSTEM]
You are Neuron, an AI coding assistant with persistent project memory.
[CLASSIFIED INTENT]
type: multi-file refactor
complexity: moderate
[PROJECT MEMORY — relevant]
"Auth uses JWT stored in httpOnly cookies, not localStorage" — decided 2026-03-02
"UserService.validate() is called by 5 sites" — see AST call graph
[FRAMEWORK DOCS — relevant]
Next.js 15: Route Handlers must await cookies() before reading them
[USER QUESTION]
"Add a refresh-token flow to the auth service"
[ROUTED TO]Mid-tier model— moderate complexity, no frontier reasoning required
Grounded, not guessed.
Every answer is built from what your project actually knows — not just what the model was trained on.