KnoLo

Local-first knowledge infrastructure for AI. Build portable .knolo packs with deterministic retrieval and optional semantic reranking—no vector DB.

[HIVE | A.I. @hiveforensics]

Knolo v4 is now live. 🧠

Today we’ve merged the next major evolution of Knolo into knolo-core: Knolo v4 / Phase 5 TrustBench.

This release pushes Knolo further toward the infrastructure we believe AI systems need: local-first knowledge that can be inspected, verified, reproduced, and trusted.

With v4, @knolo/core now builds the new Knolo pack format by default.

What’s new:

🔐 Verifiable .knolo packs
Strict v4 structure, section validation, cryptographic digests, source manifests, and fail-closed corruption detection.

🧾 Retrieval receipts
Queries can generate verifiable receipts binding retrieval results to the query, mounted knowledge state, evidence, and retrieval plan.

📍 Exact evidence spans
Results can point back to the precise source material used during retrieval rather than returning an opaque similarity result.

🧭 Deterministic retrieval plans
Knolo can expose and hash the retrieval plan behind a query, making retrieval behavior reproducible and auditable.

⚖️ Answer / clarify / abstain decisions
Retrieval now has an explicit decision layer, including the ability to abstain when the requested evidence scope is empty.

🧪 TrustBench
We added a reproducible conformance suite covering deterministic IDs, ranking scores, retrieval-plan hashes, receipts, Recall@K, MRR, nDCG, abstention precision, and corruption rejection.

🛠️ A much stronger CLI

knolo inspect
knolo verify
knolo migrate
knolo query --receipt
knolo explain
knolo diff

Legacy v1–v3 packs remain readable, while TypeScript becomes the reference implementation for the v4 contract.

This is an important step in what we’re building with Knolo.

AI systems increasingly depend on external knowledge, but retrieving information is only part of the problem.

We also need to be able to answer:

What knowledge was mounted?
What evidence was retrieved?
What retrieval process produced it?
Has that evidence changed?
Can another runtime reproduce the result?
Can we verify it independently?

Knolo v4 starts making those questions part of the knowledge layer itself.

Local-first.
Deterministic.
Portable.
Verifiable.

And open source.

@knolohq

[KnoLo @knolohq]

Knolo’s DFINITY ICP Canister Adapter is now live on `knolo-core`.

Deterministic `.knolo` pack retrieval directly on-chain.

No vectors. No cloud. No middleware.

Rust adapter + CLI + working local `dfx` e2e.

https://t.co/5M87cHkgQe

[KnoLo @knolohq]

Introducing **Knolo Agents** — an open-source Rust runtime and TypeScript SDK for building reliable, inspectable AI agents.

Typed execution graphs. Explicit authority. Policy enforcement. Deterministic replay.

Built for governed agent workflows.

https://t.co/E9uiWQF01o

[KnoLo @knolohq]

AI knowledge shouldn’t require a cloud service, a vector database, or permanent infrastructure.

Knolo packages knowledge into portable, local-first `.knolo` files with fast, deterministic retrieval and optional semantic reranking.

This is Knolo.
https://t.co/LAjX7c7zVY