neutron AIProve the change before moving your data
Keep the JSON store, Redis history, pgvector retrieval or other memory service you already use while evaluating one useful difference. Use invented support preferences first. This is a reproducible functional check, not a benchmark or a claim that Neutron AI outperforms another product.
Run the supplied check
Complete account, Nucleus and token setup. Download evaluate-memory.mjs, inspect it, then run it in Node.js 22.18+ from your backend environment:
node evaluate-memory.mjs
It reads NEUTRON_API_TOKEN, NEUTRON_NUCLEUS_ID and optional NEUTRON_API_URL. It creates two uniquely named synthetic Scopes, writes two preferences, checks each requested Scope's recall, builds a 400-token context request, forgets one record and checks fresh recall and context. It finally tombstones its remaining fixture. No package install or model-provider key is needed.
A successful result has passed: true, six named checks and billableMemoryRequests: 7. The normal path makes nine HTTP requests, including two unbilled forget requests. The runner stops after at most twelve requests, uses per-request timeouts, does not follow redirects with credentials and prints no tokens or memory content. A failure exits nonzero. If the process or network is interrupted, inspect the evaluation: Scopes and clean up remaining fixtures before using the Nucleus for anything else.
The Scope check verifies retrieval selection. It does not prove that a malicious end user cannot choose another Scope through your application. Test your backend authorisation separately, and use a second Nucleus/token for a cross-Nucleus rejection test. The runner does not prove asynchronous archive/queue deletion; follow security evaluation for that evidence.
Record the same workload in your baseline
| Question | Existing implementation | Neutron AI evaluation |
|---|---|---|
| Does the next task retrieve the right customer's preference? | Record observed IDs and summaries | Record observed IDs and summaries |
| Does a correction or revoked consent stay out of future context? | Repeat recall and your cache paths | Forget, then repeat recall and fresh context |
| What implementation work is required? | Record code and operational steps | Record backend auth, Scope mapping, calls and failure handling |
| What does the model actually receive? | Save a sanitised prompt fixture | Save a sanitised promptContext fixture |
| What does the workload cost? | Your storage, retrieval and model charges | Count Neutron requests plus your unchanged model charges |
| What improves the user's task? | Measure repeated questions or stale answers | Run the identical prompts and review both outputs |
Run with the same authorised records, task prompts and model configuration. Record actual latency, token counts, errors and operator time if you measure them; do not treat a token estimate as guaranteed savings. Do not migrate because a synthetic check passes. Keep a rollback path and choose a dependency only when the measured improvement justifies its maintenance.
For sources and conflicting claims, use the knowledge/decision guide. For decision review, start with the Free support-escalation workflow rather than an industry pack.