Introduction
AI proposes. Mnemix governs.
Mnemix is the memory and enrichment layer for AI agents — and it is built toward contextual intelligence at scale: a persistent, evidence-cited, governance-gated, self-correcting layer that any agent plugs into and stops drifting. It is not a chatbot with memory bolted on, not a vector database pretending to be cognition, and not a RAG pipeline. It is the layer that sits between an agent's proposed action and the world.
Memory is the foundation. Governance is the moat.
The problem Mnemix solves
Every AI agent today drifts. After a long context window, a coding agent re-derives facts you established hours ago. A voice agent forgets a returning caller. A finance agent claims a balance the data doesn't support. A support agent quotes a different number on turn 40 than it confirmed on turn 5.
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. Mnemix's answer is different:
- Never miss anything. Every fact carries provenance. Every locked value carries evidence. Every decision is rerunnable. A contradiction bounds an old fact — it never silently deletes it. Forgetting is failure.
- At scale. Tiered retrieval under a voice-grade budget — designed for sub-300ms voice recall — tenant-isolated by row-level security, with audit logs and deterministic decisions.
- Governance, not memory. Mnemix is building toward the gate that won't let an agent act until invariants are satisfied — the locked fact that can't be dropped across a 40-turn session, the policy bundle that codifies a domain's rules and replays them on every proposed action, the feedback loop that learns from outcomes.
What works today vs. what's in beta
Mnemix 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} |
| Validation gate | In development (beta roadmap) | Spec-frozen contract, not yet publicly callable |
Everything runs on one base URL: https://mcp.mnemix.ai (REST and the MCP transport at /mcp).
What "AI proposes, Mnemix governs" means in practice
An agent never acts directly on a high-stakes decision. It proposes an action. Mnemix evaluates that proposal against the active policy bundle and the relevant locked facts, then returns a verdict:
| Verdict | What happens |
|---|---|
allowed | The action proceeds. |
denied | The action is blocked and logged (a hard barrier in high-stakes mode). |
needs_human | The action is queued for human review. |
Every verdict ships with the evidence refs that justify it, and every verdict is rerunnable — an auditor can re-execute any past decision against the exact rules and facts that were active at that moment. That rerunnability is what wins enterprise procurement, and it's something a memory API fundamentally cannot do.
The six substrate primitives
Everything in Mnemix is built from six stable primitives. Implementations (embedders, rerankers, storage tiers) evolve; these bricks do not.
| Primitive | What it is |
|---|---|
| locked_facts | Confirmed values with a supersession chain, each citing evidence. |
| evidence_refs | Source pointers with content hash and bi-temporal validity — provenance for every claim. |
| policy_bundles | Versioned, rerun-stable rule definitions for a domain. |
| validation_gate | The deterministic checker that produces verdicts. |
| feedback_loop | The loop that learns from real outcomes. |
| bi_temporal_validity | The two-timeline model that makes 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
Mnemix never hardcodes a domain. The same gate design validates a finance reconciliation action and a coding API-contract proposal — only the bundle differs.
| Domain | What Mnemix prevents |
|---|---|
| Finance recon | An agent claiming a balance the data doesn't support |
| Coding agents | Drift on package versions, test runners, or API contracts after long context |
| Customer support | A different number on turn 40 than confirmed on turn 5 |
| Voice agents | Failing to recognize a returning caller or losing the thread mid-call |
Where to go next
- Installation — connect the CLI, the MCP server, and the SDK.
- Quickstart — your first live recall in five minutes.
- Core concepts — the six primitives, in depth (see the sidebar).