Meraki FAQ

Meraki is an iterative layer for AI agents. It turns explicit corrections and outcomes into small proposed guidance changes, applies approved changes only to relevant work, and preserves a way to inspect and undo them.

It is the layer that connects one run to the next through evidence and reviewable change. The model still generates the answer. Meraki keeps the correction, proposes what might change, controls where that change applies, and compares a later result.

No. Meraki is not a foundation model and does not replace an agent framework. It is designed to sit beside models and agents as an evidence, guidance, evaluation, and rollback layer.

No. Memory and retrieval help recall information. Meraki focuses on whether evidence should become guidance, which tasks may receive it, whether it improved a later result, and how to reverse it.

Explicit corrections, selected outputs, references, choices, and observed outcomes can become evidence when their source, task, and time remain attached. Model inference alone is not durable permission.

No. A correction can support a candidate change, but candidates do nothing until a person or explicit policy approves their wording and scope.

Meraki Core can model facts, current state, goals, identity declarations, behavior, cognitive patterns, communication, voice, taste, judgment, workflow, examples, anti-patterns, modes, and uncertainty. Retrieval selects only what the current task is allowed to receive.

They answer different questions. A facet says what kind of guidance something is. A horizon says how long it should matter. A temporary goal should not be treated like a durable identity claim.

The Core design compares a baseline, raw memory, the scoped MerakiPack, and a targeted-removal run without the guidance being tested. Independent evaluation at meaningful scale is still unfinished.

Approved guidance carries explicit scope, project, task-family, mode, horizon, and sensitivity boundaries. Pack construction is deterministic so an unexpected match can be reproduced and repaired.

That is a core requirement. Accepted changes keep evidence, scope, profile version, and outcome records. They can be weakened, respecified, replaced, revoked, or rolled back.

Not by design. An agent can help record evidence or propose a candidate. Durable activation requires a person or an explicit approval policy, and untrusted retrieved text cannot promote itself.

The independent Meraki Core repository implements and tests a local deterministic evidence-to-rollback loop and connected runtime paths. Hosted invited-user proof, live PostgreSQL and pgvector proof, and independent human evaluation remain open.

Meraki is preparing an invited beta. Join the waitlist for access updates. There is no public release date or pricing claim yet.