AKF: portable knowledge with evidence.
The Agentic Knowledge Format turns curated knowledge into an artifact you own: readable, versioned and authenticated, with provenance and selected supporting evidence. Keep what your agents have established when you change model, provider or machine.
The revolution is what you can take with you.
This lifecycle illustration assumes an existing valid signed bundle and independent recipient trust. It is not a from-zero bundle-authoring tutorial.
Imagine a project decision: “Use a single writer for the operational store.” Its curated page explains why, cites a source revision and a selected test report, and declares the supporting evidence. Another agent can inspect the reason, not just repeat a conclusion. This is an illustrative decision, not a benchmark result.
| Bundle part | What survives the transfer |
|---|---|
| Markdown pages | Readable decisions, procedures and reusable knowledge |
| KDL sidecars | Identity, revision, provenance, authority and disclosure |
| Declared sources and selected evidence | The support and dependencies needed to interpret claims |
| Retained history | Earlier revisions and supersession |
| Manifest and signatures | Payload membership and bytes authenticated with trusted keys |
- CurateConclusion + evidence
- PublishValidate + authorize + sign
- TransferRecipient verifies its own trust
- RetrieveBounded pages + exact citations
Readable bundles need no database. Authenticated retrieval needs a compatible owner and independently configured recipient policy, trusted keys and manifest pin. Copying the bundle grants no publication or execution rights. Signatures authenticate the snapshot; factual truth, freshness and completeness still require checking.
One knowledge owner. Many agent sessions.
JARVIS/Mnemos authenticates snapshots, enforces project/disclosure
policy, and controls publication through ASM, the Agentic State Machine.
PRISM validates the wire and forwards requests through
prism-mem knowledge, delivering whole pages with citations and omissions.
It runs no second writable memory engine.
Raw chats, current tasks and live capabilities remain operational Context. Durable conclusions become AKF Knowledge through deliberate curation and authorized publication. Provider startup is an explicit binding: configuration alone does not prove that a provider trusted the hook or delivered the packet.
Configure the adapter and startup binding · Exact JSON protocol and limits
Inspect the portability evidence.
The real-owner process acceptance fixture exports a signed revision with source evidence/history, removes the original store and configuration, then retrieves using recipient-only policy. It also exercises wrong-project denial, missing/revoked trust, incorrect manifest pins and payload tampering. This checks portability independently of the original chat store; it does not prove every curated assertion or universal agent compatibility.
Read the executable acceptance fixture · Inspect the signed bundle construction
This is existing source evidence, not a newly measured benchmark. Check installed command help and the documented owner/provider requirements before use.
How determinism works.
A model proposes work. Fixed schemas, rules, policy and recorded workflow steps govern execution. The implemented doctrine gate uses no model call:
- CaptureFresh guidance and input identity
- CompareExplicitly selected baseline
- AssertComplete evidence; no added errors
- HaltRecorded result, receipts and audit
| Situation | Defined behavior |
|---|---|
| Same captured inputs and evaluation identity | Repeatable rule evaluation without model sampling |
| Fresh invocation | New IDs and observation; changed files can change the verdict |
| Added error against the accepted baseline | Gate assertion fails |
| Incomplete/incompatible evidence | Cannot qualify as a passing gate |
| Completed-gate replay | Historical result under current admission; no fresh filesystem check |
| Interrupted-gate recovery | Original intent retained; unresolved operations stop without blind retry |
PRISM supplies the evaluation and comparison. JARVIS admits the tool under policy and records outcomes. A pass means no added errors against the selected baseline; existing errors and warnings may remain. It is not proof that the repository is safe. Model reasoning, timestamps, IDs and external tool effects are not claimed to be deterministic.
AKF uses schema, manifest, signature, scope and revision checks before Knowledge delivery; ASM separately requires publication authority. Signed content remains reference material and cannot override runtime policy.
Run the capture and comparison example · Configure the JARVIS gate and durable recovery
Gate/recovery requires the JARVIS unstable builds identified in that
tutorial and the documented Linux capture support. Website publication does not
release the runtime, install hooks, publish a brain or activate a scheduler.