Developer docs

Quickstart

From zero to scored memory in three conceptual steps. KAPEX is middleware — your app calls KAPEX, KAPEX returns context, you inject it into any LLM.

1. Get a pilot key

Start the pilot and we'll send you a key plus the configured SDK and MCP packages. Pilot keys are scoped, rate-limited, and revocable.

2. Send a conversation turn

Pass the user's message and your LLM's reply to KAPEX through whichever surface you prefer (SDK, MCP, or REST). KAPEX scores the turn against 12 signal dimensions, updates the memory graph, and applies the decay model.

3. Inject memory into your LLM

Request a salient-memory context block at query time and inject it into your system prompt. The model now references prior turns — without you changing the model. Provider-agnostic.

import kapex

# 1. Send the message to KAPEX
memory_context = kapex.process(user_id, message, session_id)

# 2. Inject memory into your prompt
enhanced_prompt = system_prompt + "\n\n" + memory_context

# 3. Call your LLM as usual
response = your_llm.chat(enhanced_prompt, message)

Integration pattern

Three steps, in order: deploy the container, connect your app, inject memory into your prompt. KAPEX runs in your infrastructure — your app talks to your KAPEX endpoint, never to ours.

Typical integration takes under one day for a senior engineer. No model changes, no prompt rewriting, no vendor lock-in. User data never leaves your VPC.

Three surfaces, one engine

Pick whichever surface fits your stack. They're functionally equivalent — the same scoring engine, decay model, retrieval, and safety pipeline runs behind all three.

Auth & rate limits

Authentication

• API keys are hashed at rest, never stored in plaintext. Rotation supported.
• Hosted MCP servers use Cognito-issued RS256 JWTs with short TTLs and rotating signing keys.
• Enterprise plans support SSO via SAML 2.0 (Okta, Azure AD, Google Workspace) and SCIM 2.0 provisioning.

Rate limits

When you hit the cap, the API returns a 429 with a Retry-After header. No overage charges. The SDK auto-retries with exponential backoff.

Errors

Standard HTTP status codes. JSON body with a stable error shape — error message, code for programmatic handling, and a request_id for support. Always include request_id when escalating.

Deployment

Two paths, depending on your data-sovereignty needs:

• KAPEX Cloud — hosted in AWS us-east-1. Default on Pilot, Build, and Scale plans.
• Customer-hosted — deploys into your AWS, GCP, or Azure VPC, or on-prem. Ships as a CloudFormation template or signed Docker package. Sandstone handles deployment. Available on Enterprise. See Deployment model.