Get started · 1 / 42
Introduction
Northsun is the memory and enrichment layer for AI agents — a live caller memory REST surface plus a beta hosted MCP for chat, coding, and workflow agents.
Every agent proposes. Northsun governs.
Northsun is the memory and enrichment layer for AI agents. Its differentiator is governed context intelligence: preserving durable context, attaching evidence and temporal validity, and making the right context available to voice, chat, coding, and workflow agents. Northsun does not build those agents; it gives them API and MCP surfaces for context and memory.
The problem Northsun solves
Agent systems fail when context rots, disappears between sessions, loses its source, or reaches the model at the wrong time. A coding agent can re-derive a decision that was already settled. A voice agent can treat a returning caller like a stranger. A workflow agent can act on a value without knowing when it was true or where it came from.
The industry's answer has been "give the agent more memory." That's the wrong layer. More memory without governance just means the agent can now confidently retrieve and act on the wrong thing. Northsun's answer is different:
- Memory and enrichment today. The frozen caller surface recalls prior interactions, adds available Trestle and Twilio Lookup context, and writes the completed call back for next time.
- Governed context as the differentiator. Evidence refs, locked facts, bi-temporal validity, policy bundles, and validation gates describe how context is designed to become auditable instead of merely retrievable.
- Salience as the spine. The goal is the right context, in the right amount, at the right time. Proactive Beacon behavior remains design-stage until its runtime and proof gates graduate.
What works today vs. what's in beta
Northsun ships in layers. Be precise about which layer you're wiring:
| Layer | Status | Surface |
|---|---|---|
| Voice memory + enrichment | Live | POST /v1/recall_and_enrich · POST /v1/calls/end · GET /v1/caller/{phone_number} |
| Authenticated MCP context tools | Beta surface | Hosted MCP at /mcp; inspect tools/list for the current contract |
| Validation gate | Built in source; not public | Outside the frozen public contract |
Everything runs on one base URL: https://mcp.northsun.ai (REST and the MCP transport at /mcp).
What governed execution means in practice
The governed execution model is: an agent proposes an action; a gate evaluates the proposal against the active policy bundle and relevant facts; the system returns a verdict. This section describes the beta architecture, not a public REST endpoint:
| Verdict | What happens |
|---|---|
allowed | The gate's verdict is that the action may proceed. |
denied | The gate's verdict is that the action is blocked. |
needs_human | The gate's verdict is that a human should decide. The gate returns and records the verdict; it provides no review queue. |
The design records the evidence and versioned state needed for reruns. Public claims about production rerun behavior remain gated on the validation surface and runtime proof.
The six-part governed architecture model
The North Star uses six concepts to describe the governed system. They do not all have the same ship state: some exist in the current substrate, some are not publicly callable, and one remains design-only.
| Concept | Current status | What it is |
|---|---|---|
| locked_facts | Built substrate; no frozen public facts API | Confirmed values with a supersession chain, each citing evidence. |
| evidence_refs | Built substrate; no frozen public evidence API | Source pointers with content hash and bi-temporal validity — provenance for governed claims. |
| policy_bundles | Built substrate; not publicly callable | Versioned rule definitions for a domain. |
| validation_gate | Built in source; outside the frozen public contract | The Stage-1 checker that produces verdicts; no public validation endpoint is promised here. |
| feedback_loop | Built in source; outside the frozen public contract | Records what served context turned out to be, and moves the confidence of the memories a caller names. recall_and_enrich exposes the context_audit_id and the memory ids feedback needs when the voice feedback loop is enabled on a deployment; the context path does not. It adjusts a stored confidence value, which ranking reads — it does not learn or tune weights. |
| bi_temporal_validity | Built for evidence_refs and locked_facts | The two-timeline model that makes governed-record history reconstructable. |
(Memory objects — the embedded memories your agent recalls — are the storage layer these primitives govern. Per a locked design decision, memory objects themselves are not bi-temporal; only evidence_refs and locked_facts carry the full two-timeline model.)
Domain-agnostic by design
Northsun does not hardcode a domain in the evaluator: domains arrive as bundles. One finance bundle exists in source; none exists yet for coding agents or customer support.
| Domain | What a bundle would check | Status |
|---|---|---|
| Finance recon | An agent claiming a balance the data doesn't support | A bundle exists in source |
| Coding agents | Drift on package versions, test runners, or API contracts after long context | Specified, not built |
| Customer support | A different number on turn 40 than confirmed on turn 5 | Intent, not built |
| Voice agents | Failing to recognize a returning caller or losing the thread mid-call | Not a gate bundle: recall_and_enrich is the production path for returning callers |
Where to go next
- Installation — connect the CLI, hosted MCP, or an HTTP client.
- Quickstart — prove one recall, submission, and later write-back.
- Authentication — keep keys and tenant scope correct.
- MCP server — connect voice, chat, coding, or workflow agents.
- Core concepts — the governed architecture model, with ship-state labels (see the sidebar).