Core Concepts

Mnemix is built from six primitives. Implementations cycle — the embedder, reranker, storage tiers, and rule engine are all swappable as the field evolves. The primitives do not break. They are the LEGO bricks; everything else is built on top.

| Primitive | One line | Status | | :--- | :--- | :--- | | locked_facts | Confirmed values with a supersession chain, each citing evidence. | Substrate live | | evidence_refs | Source pointers with content hash and bi-temporal validity. | Substrate live | | policy_bundles | Versioned, rerun-stable rule definitions for a domain. | Beta | | validation_gate | The deterministic checker that produces verdicts. | In development | | feedback_loop | The loop that learns from real outcomes. | Beta (endpoint live) | | bi_temporal_validity | The two-timeline model that makes history reconstructable. | Substrate live |

A note on memory objects — the embedded memories your agent recalls across storage tiers. They're the storage construct the primitives govern, not a primitive themselves. Per a locked design decision (D3), memory objects are not bi-temporal: only evidence_refs and locked_facts carry the full two-timeline model. An as-of reconstruction query answers from those two tables; memory objects record when they were ingested, nothing more.

How facts are addressed

locked_facts are addressed by entity + key — entity_id (who or what the fact is about) and fact_key (which fact). evidence_refs are deliberately address-agnostic: provenance rows carry source, hash, trust priority, and time — they're linked from facts, never looked up by domain key. Tenant isolation is enforced on every row by row-level security.

How they fit together

evidence_refs  ──cited by──▶  locked_facts  ──read by──▶  policy_bundles
                                                                │
                                                          evaluated by
                                                                ▼
proposed action  ───────────────────────────────────▶  validation_gate ──▶ verdict
                                                                │           (allowed | denied | needs_human)
                                                          outcome flows to
                                                                ▼
                                                          feedback_loop ──▶ adjusts memory confidence,
                                                                            audits the context read

Two cross-cutting ideas run through all six:

• Bi-temporal validity — facts are never deleted, only bounded in time. Mnemix can always answer both "what is true now?" and "what did the agent believe was true last Tuesday at 2 PM?" — for evidence and locked facts.
• Determinism — the same context as of the same moment produces the same verdict. That's what makes decisions auditable rather than approximate.

Read them in order, or jump to whichever primitive you're wiring up first.