← Blog

· Hive AI

Knolo Research: Three Open Questions on Replay, Persistence and Authority

The Knolo research index names three open questions: replaying retrieval, persisting knowledge by content address, and portable authority. What Knolo documents today.

Knolo is the open-source, local-first Knowledge Image runtime and format, released under Apache-2.0 as @knolo/core on npm and developed in the open in the Knolo GitHub repo. The Knolo research index on knolo.org sets out three questions the research program cares about: can a result be replayed, where did the context come from, and what can an agent do.

The index is careful about status. Its three records, KNF-RP-001, KNF-TR-001 and KNF-RFC-001, are labeled example records. They are not published papers, they carry no authors or abstracts, and they say "Do not cite." The page puts it plainly: unpublished is better than invented. This post follows the same rule. It explains the three questions and points to what Knolo already documents today, without claiming results that have not been published.

Knolo question one: can the result be replayed?

KNF-RP-001, Deterministic retrieval and replay, frames the question as: what inputs, encodings and runtime bounds are required for a retrieval result to be replayed across environments?

Knolo's current answer starts with the image itself. A Knolo Knowledge Image is pinned by digest, and a Knolo query receipt binds a plan root and a result root to one committed Knowledge Image state. The Knolo query receipts docs and Knolo verification docs describe what is checked today. A receipt is an integrity check. It records what was read; it does not prove that a claim is true. The open research question is what else must be pinned, beyond the image, for a replay to match on another machine.

Knolo question two: how should knowledge persist?

KNF-TR-001, Content-addressed knowledge persistence, asks how a Knowledge Image should persist documents, roots and history so another machine can verify the same bytes later.

The Knolo storage docs describe today's V5 behavior. Storage is a single committed image with snapshot readers. Object and event segments are immutable, and committed bytes are replaced with a temp file, fsync and rename rather than edited in place. An optional writer lease is a local coordination primitive, not a distributed consensus protocol, and multi-writer replication across hosts is not a V5 storage mode. Those limits are stated in the docs, and they mark where the research question begins.

Knolo question three: what can an agent do?

KNF-RFC-001, Portable authority boundaries, asks how policy, identity and delegated authority should travel with a Knowledge Image without implying a completed distributed database.

The Knolo policy and authority docs show the current line. V5 policy evaluation is bound to the committed policy root, and signed authority envelopes, bounded delegation, key rotation and keyring-root selection are specified in KIP-0008 through KIP-0011. The host still owns credentials, the decision to invoke a mutation and every external effect. Knolo records allow or deny; it does not act on its own.

How Knolo research becomes a public record

The research index says future documents will be published with identifiers, authors, dates, abstracts, status, download links and citation information. Until a record reaches that state, it stays marked as an example. Changes to the format itself go through the Knolo Improvement Proposals (KIPs) process, and the Knolo open technology lifecycle page explains the explicit lifecycle states that tell people what they can rely on.

Contribute to Knolo research

If one of these questions overlaps your own work on deterministic retrieval, provenance or agent authority, the Knolo community page explains how to get involved. Developers can start from the Knolo developer docs on knolo.dev and the Knolo core source on GitHub.