Memory layers should answer, not hallucinate

Agent memory is only useful when the next consumer can tell the difference between remembered fact, current inference, and missing context. The common failure mode is a recall layer that emits fluent prose even when the underlying record is partial or stale.

That turns a data problem into a trust problem.

At labs21, we bias toward memory systems that preserve the boundary between retrieval and interpretation. A recall surface should return enough structure that an agent can inspect the source, attach confidence, and decide whether to continue or ask for more evidence.

await memory.write({ scope: "agent" });

async function resolve(query: string) {
	return await recall(query);
}

The implementation detail matters less than the contract. When memory cannot answer, it should fail loudly and specifically. A plausible lie is the worst possible cache hit.