FAQ

Is SPM open source or self-hosted? No. SPM is a hosted service accessed through the provider proxy or MCP. It requires no customer-side deployment.

Do I have to move my models? No. Models remain in your provider accounts (DeepSeek, OpenAI, Azure, Anthropic, Gemini, or any compatible endpoint). SPM forwards requests in each provider's native dialect.

Does SPM add LLM calls to my requests? No. Recall, ranking, admission, and compression are deterministic. Provider charges cover inference only; memory growth does not add tokens.

What happens when there is no relevant memory? The evidence gate fails closed and returns an explicit UNKNOWN with a machine-readable gate reason.

Can I delete my data? Yes, per source or per tenant, through /v2/purge or the console. Each purge returns an HMAC-SHA256 signed receipt stored with the purge record and verifiable by the operator. Receipts include their key id and remain valid after key rotation. Two caveats remain: records created before signed receipts are unsigned, and pre-purge backups age out on their own schedule. Cryptographic erasure through per-tenant key destruction is on the roadmap.

Can another tenant ever see my memory? No. Every object carries a tenant id. Row-level security, namespace fences, signed tokens with replay protection, and a final evidence-gate recheck at render time enforce isolation.

What providers does the proxy speak? OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages, each without cross-dialect rewriting.

What if the console is down? Traffic continues because the request plane does not depend on the console plane.

Where are the numbers from? The Benchmarks page provides the scorecard, evidence annotations, and methodology.