Synapse is managed pgvector with tenant-scoped retrieval, a real embedding pipeline, and domain indexers. Long-term memory for Cortex agents and any external system that wants the same shape.
Waitlist opens with launch. Check back when the page flips to Available.
Agents forget context. Running pgvector plus an embedding pipeline plus tenant scoping plus re-embedding on model upgrade is operational work that pulls focus away from the product. Synapse takes the operational work and leaves the product to you.
Every embedding stored with mandatory (tenantId, contextDomain). Every query requires both. No defaults, no ambient context. Cross-tenant queries are a separate admin API.
Ingest, chunk, embed, store. Re-embedding scheduled per tenant when the embedding model upgrades; old vectors retired in the same pass.
Vector similarity plus keyword filters. Top-k chunks return with source pointers so the agent prompt can cite, not just answer.
Synapse exposes retrieval as a Cortex tool. Tenant scope is injected at the tool invocation layer; the agent prompt cannot specify or override it.
Hosted vector stores you wire into your own retrieval pipeline. Tenant isolation is your job. Synapse ships the isolation as a primitive.
Synapse is built to be called from Cortex agents with attribution that survives tenant audits. The integration is the point, not an extra.
Cortex agents query Synapse via tool. The tool invocation injects the run's tenant scope before Synapse sees the query, so the agent's prompt cannot exfiltrate cross-tenant embeddings. Cortex is the agent runtime that ships the integration.
Synapse is not available yet. Leave your email and we'll let you know when it launches.
