Architecture Overview

Synapse Layer is a multi-tenant memory infrastructure with three access layers, a 3-stage security pipeline, and persistent vector storage.

On this page

Access Layers

Agents and applications can interact with Synapse Layer through three complementary interfaces:

REST API v1

16 endpoints under /api/v1/ for programmatic access. Includes memory capture, recall, token management, usage metrics, handover, and consent management. Authenticated via Connect Tokens.

MCP Server (Streamable HTTP)

13 tools exposed via the Model Context Protocol at /api/mcp. Compatible with Claude Desktop, Cursor, and any MCP-compliant client. Uses streamable HTTP transport.

SDKs (Python & TypeScript)

High-level client libraries that wrap the REST API with built-in sanitization, privacy, and intent validation. The Python SDK is the reference implementation; the TypeScript SDK provides equivalent core operations.

Security Pipeline

Every memory operation passes through a 3-layer security pipeline. These layers are independent and non-bypassable (Intent Validation is always active; the other two are enabled by default):

Layer 1

Semantic Privacy Guard™

PII detection and redaction using 12+ regex pattern categories. SHA-256 forensic hashes for audit.

Layer 2

Differential Privacy

Calibrated Gaussian noise on embeddings with configurable ε budget. L2-normalized after injection.

Layer 3

Intelligent Intent Validation™ v2.1

Intent classification with confidence gating and critical keyword auto-promotion. Always active.

Storage Engine

Synapse Layer uses PostgreSQL with the pgvector extension for persistent storage:

  • Vector storage — 1536-dimensional embeddings indexed via HNSW (m=16, ef_construction=128)
  • Encryption at rest — AES-256-GCM with per-operation random IV and GCM auth tag validation
  • Structured data — Intent classification, Trust Quotient, timestamps, and metadata in relational columns
  • Audit trail — ForgeEvent and CognitiveAudit tables for immutable operation logging

Multi-Tenant Model

Tenancy is enforced at the data layer:

  • 1 user = 1 tenant = 1 isolated memory space
  • tenantId is a NOT NULL foreign key on every memory record
  • Tenant resolution is derived from authentication, never from request parameters
  • Cross-tenant queries are structurally impossible at the database level

Note

Within a tenant, multiple agents can share memories via cross-agent recall. Between tenants, zero data sharing occurs.

Deployment Model

Synapse Forge is the hosted production environment. It runs as a single application serving the REST API, MCP server, dashboard, and documentation. The Python SDK can also run locally or be embedded in any Python application.

Next Steps

Storage Layer

Deep dive into pgvector, HNSW indexing, and encryption at rest.

Security Pipeline

Detailed walkthrough of the 3-layer security pipeline.