Skip to content

Quickstart: Initialize a Project and Record Your First Memory

Run this from your project’s root directory:

Terminal window
neuron init

neuron init detects every coding-agent harness present (.claude/, .codex/, .github/, .cursor/, or a bare AGENTS.md, defaulting to creating AGENTS.md if none is found), pre-downloads the local ONNX models neuron runs offline, and wires a recall hook into every detected harness that supports one. It also writes a starter neuron.yaml and appends a ## Memory Store Protocol block to each harness’s instructions file.

You can also just tell your coding agent “Set up neuron memory for this project” — it runs the same interview and configures the project for you.

Terminal window
neuron memory add --category learning "Always use the Repository Pattern for database access in src/services"

This writes one entry to .neuron/learning.md, plain markdown with a small YAML frontmatter block (id, createdAt, importance, tags). Nothing here is a database row you’d need a viewer to inspect — open the file in any editor.

Terminal window
neuron memory query "How do we handle database access?"

This runs hybrid search — semantic and keyword matching combined — over every category configured for query, and returns the entry you just wrote even though the query text shares no exact phrase with it.

Recall — neuron surfacing a relevant memory back to your agent — is solved by the hook neuron init just wired. It does not solve the other half: getting your agent to write to the store in the first place. An A/B test measured that gap directly — under realistic multi-step conditions, an agent with no nudge recorded a fix only 20% of the time it should have. To close it, neuron init also wires a pre-stop hook that, once per session, forces one more turn with a reminder if nothing has been recorded yet. See the harness adapter pages for which harnesses support this mechanism and how.

Source: README.md “Write-side compliance gets a nudge, not just a reminder”.