stop starting from zero.
retain useful facts & where they came from in a vault controlled by the user.
the context layer.
useful context for your AI. control stays with you.
your preferences, notes & decisions help an agent do its job. the context layer proposes a way to share only what a task needs, with a clear purpose, a time limit & a record of what happened.
remember more.
reveal less.
switching tools can mean explaining yourself again, or handing over far more history than a task needs. useful memory should come with boundaries.
share for one purpose.
check the task, recipient & permissions before releasing a small, time-limited bundle.
see what happened.
keep a minimized receipt of the decision & action. review proposed memory changes separately.
prepare for a meeting.
an illustrative flow with fictional notes. the interactive demo uses synthetic data too.
‘help me prepare for the project meeting.’
ask for the relevant notes.
the agent names the project, the meeting-prep purpose & the facts it needs.
check before sharing.
policy can allow, reduce, ask for approval or deny. unrelated personal notes stay in the vault.
send only the approved context.
the agent receives a bundle for this task & recipient, with an expiry. a receipt records the exchange.
what does that mean in the protocol?
a ContextRequest declares the purpose. policy produces a decision. a ScopedContextBundle carries only approved claims. capabilities limit the permitted action; a receipt records the result. any new memory is a proposal to review.
one boundary. six recorded steps.
the six steps below explain the full flow. open any step for detail. raw vault access never crosses the boundary.
01capturepreserve origin
record the source event, time, identity & payload digest before interpretation.
02normalizederive with provenance
derive structured claims while retaining source references, timestamps & confidence.
03vaultretain authority
keep encrypted originals & policy state inside user-controlled authority. raw data never enters bundles or receipts.
04decideallow, reduce, approve or deny
evaluate purpose, recipient, selectors, actions & expiry before granting any context.
05bundleminimum useful context
issue recipient-bound, expiring, single-use context with restrictions & provenance.
06actexecute & receipt
verify capabilities, execute the permitted action & append a minimized receipt. memory writes remain proposals.
receipt rail
request → decision → disclosure → action → writeback
a working draft.
a testable starting point.
the context layer is experimental. the public reference work & new adapter research have different boundaries.
local core + proof
the published proof covers v0.2 contract validation, an encrypted local vault, four-state policy, scoped bundles, anchored receipts, a files adapter & a local-agent consumer.
production key custody & end-to-end vendor interoperability are not established by this proof.
inspect the public starter →Poppy + context layer
an offline prototype explores Poppy draft 0.1 discovery validation & text-only conversation request preparation, with PCP consent checks & separate context layer disclosure gates.
requests are prepared, not sent. live sign-in, session authentication, business actions & external interoperability remain unimplemented.
read the scope & limits →choose your next step.
01essay
understand the idea & why it matters
02architecturesee how data moves & where authority stays
03specificationinspect the objects, rules & conformance requirements
04implementationexplore adapter profiles & implementation limits
05codetested v0.2 contracts & a local proof profile
06demorun one synthetic request through policy, disclosure & receipt