neutron AI
Changelog

Testing

Use a sandbox Nucleus and test token to validate your Neutron AI integration before enabling it for production users.

API Smoke Test

curl -s "$NEUTRON_API_URL/v1/memory/remember" \
  -H "Authorization: Bearer $NEUTRON_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<EOF
{
    "nucleusId": "$NEUTRON_NUCLEUS_ID",
    "scopeId": "$NEUTRON_SCOPE_ID",
    "type": "preference",
    "text": "Prefers concise integration examples"
}
EOF

Then verify recall:

curl -s "$NEUTRON_API_URL/v1/memory/recall" \
  -H "Authorization: Bearer $NEUTRON_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d @- <<EOF
{
    "nucleusId": "$NEUTRON_NUCLEUS_ID",
    "scopeId": "$NEUTRON_SCOPE_ID",
    "query": "integration examples"
}
EOF

SDK Smoke Test

Test these flows in your preferred SDK:

  • health check
  • remember
  • recall
  • agent context
  • reflect
  • forget
  • error handling for invalid token
  • error handling for unauthorized scope

For TypeScript knowledge and decision integrations, also test:

  • entity and relationship version creation
  • current and historical graph queries
  • bounded Context Engine output and retrieval traces
  • agent denial for an unrequested or private Scope
  • hard constraints and Pareto-efficient decision options
  • immutable decision-time knowledge snapshots
  • predicted-versus-observed outcome errors
  • review-gated learning activation
  • simulation and counterfactual separation

CLI Smoke Test

neutron health
neutron remember --scope $NEUTRON_SCOPE_ID --type preference --text "Prefers short examples"
neutron recall --scope $NEUTRON_SCOPE_ID --query "short examples"
neutron context --scope $NEUTRON_SCOPE_ID --task "Write onboarding copy"

MCP Smoke Test

Validate that your host can:

  • connect to the Neutron MCP endpoint
  • list allowed tools
  • call memory_agent_context
  • call memory_recall
  • call neutron_entity_get, neutron_graph_query, and neutron_context_build
  • call decision, consequence, outcome, or simulation tools only when the host is authorised for those workflows
  • reject write tools when the token is read-only

Product-Proof Suite

Repository contributors can run the isolated Cloudflare runtime suite:

pnpm test:product-proof

It executes 16 product journeys through the Worker, real SQLite-backed Durable Object bindings, REST, and MCP without remote or billable Cloudflare resources. The cases prove 20 capabilities across temporal graph history, contradictions, hybrid context, permissions, decisions, consequences, snapshots, outcomes, reviewed learning, simulations, provider independence, tenant isolation, audit persistence, and process eviction.

Production Readiness

Before launch, confirm:

  • tokens are server-side only
  • Nucleus and Scope IDs are authorized by your app
  • logs do not contain raw memory content by default
  • delete controls call forget
  • repeated tasks use an intentional cache policy
  • historical reads preserve both valid-time and knowledge-time semantics
  • context retrieval enforces persisted agent Scope permissions
  • predictions, simulations, counterfactuals, and observations remain distinct
  • sensitive-domain learning cannot enter context before review
  • SDK and CLI versions are pinned in your release workflow

Release notes