# Choose exact evidence for each agent

Prepare a versioned evidence selection for an AgentCore task. Capture, manifests and bounded assembly are separate from Memory recall and model submission.

Status: Implemented in source. Prelaunch; not a live service guarantee.
Evidence capture, versioned manifests, hide/restore and bounded assembly are implemented in source and locally tested with simulated storage. Public hosting, managed AgentCore wiring, automatic ranking, provider-token budgeting and live model-consumption qualification are not established.

Reviewed: 2026-10-14.
Backend source: e16b81b4a637acec91f3aed8a71cf5b0f42f0c19.
Native toolkit: merged in PR #176, reviewed at e16b81b4a637acec91f3aed8a71cf5b0f42f0c19.
HTML: /docs/context

## In short

Prepare the exact evidence for an AgentCore task separately from Memory recall. A versioned manifest selects retained inputs and how to represent them; bounded assembly checks required evidence and enforces item and byte limits. Your application controls submission to the model.

**Available today:** Capture, manifests, hide/restore and bounded assembly are implemented in source. Managed runtime wiring, automatic ranking and qualified live model use remain separate.

> Choose exact evidence for an AgentCore task while keeping recall and submission separate.

## Prepare evidence independently of recall

- **Keep required inputs** — Select the exact evidence your AgentCore task needs.
- **Bound optional inputs** — Name omitted or missing evidence and enforce item and byte limits.
- **Own the request boundary** — Your application submits the assembled representation; managed runtime interception is planned.

## Keep the evidence and its selection separate

Your application captures the bytes received at a controlled boundary. A consumer’s manifest names exact evidence and chooses full, descriptor or excluded representation. One captured record can appear differently in separate consumers’ selections.

A manifest is an ordered, numbered version. Hide and restore create another version; earlier retained manifests still name the evidence they selected. A source digest identifies the captured bytes. It does not verify the source’s accuracy or prove that a model consumed the evidence.

## Capture, select, assemble and inspect

1. **Capture exact bytes** — captureContextEvidence retains evidence under an immutable identity and digest. Changed evidence needs a new identity.

2. **Choose a versioned selection** — configureContextManifest selects up to 16 ordered entries. hideContextEvidence and restoreContextEvidence preserve prior retained selections.

3. **Check required evidence and limits** — assembleContext reads at most 32 exact keys, with a declared budget up to 16 items and 192 KiB. Unavailable or expired required evidence blocks assembly of a complete package.

4. **Inspect without hydrating payloads** — inspectContextManifest returns metadata for the exact selection. Assemble and inspect recheck authority after asynchronous reads before serving.

## Assembly is not proof that a model used it

The default response reports submission, invocation completion and business acceptance as not_observed. A controlled request adapter can report only the actual injected submission boundary; it does not intercept calls made elsewhere.

The item and byte budget describes captured evidence, not provider tokens or a complete prompt. Automatic ranking, learned compression, provider prompt caching and managed runtime wiring remain separate future integrations. No paid model is invoked by the local qualification.