Research
Recall Engine7 min read

Strong retrieval below. Strict governance above.

Aionis uses stronger candidate retrieval to recover active execution state, then keeps the same admission boundary before anything reaches the agent.

The optional Zvec ANN backend improves candidate generation without changing Aionis' core governance contract: retrieval proposes, admission decides.

The bottleneck was candidate generation

Aionis' governance layer can only adjudicate memories that reach it as candidates. The strict ID-neutral MGBench diagnostic showed that safety stayed intact, but active-route promotion could drop when the bounded SQLite scan did not surface enough current-state evidence.

That is a retrieval-engine problem, not a reason to weaken the admission gate. The right fix is to improve candidate generation while preserving the same use_now, inspect_before_use, do_not_use, and rehydrate contract.

Why Zvec fits the local Runtime

The Zvec backend is optional and local. It is a sidecar ANN index used for semantic candidates; it does not become the source of truth. Memory rows, lifecycle state, authority, feedback, and traces still live in the Runtime stores.

That separation matters. Aionis should not become a vector database wrapper. It should use vector search as one candidate source, then apply runtime-owned governance before the agent sees memory.

Strict ID-neutral result

On the 40-scenario MGBench strict interference holdout, Aionis observe->guide with Zvec ANN reached 40/40 product-positive cases, 100% active recovery, 0/40 unsafe direct-use, 100% rehydrate recall, and 100% trace coverage. The average context size was 1,306 characters.

A fresh rerun with a new Runtime, new SQLite stores, new Zvec sidecar, and new scope prefix reproduced the same outcome: 40/40 product-positive cases, 0/40 unsafe direct-use, 100% rehydrate recall, 100% trace coverage, and 1,305 average context characters.

What the result means

The result supports a precise product claim: stronger candidate retrieval can improve active-state recovery without weakening memory governance. It does not claim patch-success superiority or model reasoning superiority.

The product architecture becomes cleaner: semantic ANN, lexical, structured, graph, and recent sources generate candidates; Aionis adjudicates whether each candidate may act, must be inspected, should be blocked, or needs rehydration.