Live architecture demo

Watch the AES Research patterns fire on your own input.

Type a question or research request below. A Cloudflare Worker dispatches to Claude with a system prompt that exercises the full stack: skeptic membrane pre-check, two-tier memory lookup against a scoped fixture, and attention routing between native reasoning and tool use. The thought-stream panel on the right shows each layer firing before the final answer.

This demo is scoped and rate-limited. It uses a throwaway fixture memory and a dedicated prompt — the real AES Research systems running in production are not exposed here. The point is to observe the architecture in motion, not to probe the implementation.

What you're seeing

Skeptic membrane — the input is checked for prompt-injection, ambiguity, and scope violations before the agent acts. Community discussion of Claude Routines explicitly flagged prompt-injection risk on triggered agents; this is the mitigation.
Two-tier memory lookup — the scoped fixture MEMORY.md index is scanned; only entries the index marks relevant are loaded into context. Watch the stream report which entries got pulled and which were skipped.
Attention routing — the policy decides whether the step is native-LLM-cheap, tool-necessary, or skill-appropriate. Each decision lands in the stream with the reasoning attached.
Response — only after the three layers pass does the agent respond. If the membrane rejects, you'll see the rejection reason rather than a response.

Honest note on what this demo's memory actually does

The memory this demo shows is composed of two pieces, and it is helpful to keep them separate.

Fixture memory — five hardcoded entries in the Worker's system prompt. The model scans these and reports which it loaded. This is a stand-in for a real MEMORY.md index; it does not grow or learn.
Per-session conversation history — your turns in this browser tab accumulate in local JavaScript memory and are sent with each new request so the agent has context within the session. It is not persisted server-side, it is not shared across tabs, and it resets when you reload or click Reset.

The production AES Research memory system — the one described in the two-tier memory writeup — persists across sessions, supports index scans plus per-topic files, and is written to atomically to survive compaction (see context survival). None of that runs here. This demo is the whiteboard sketch of the architecture, not the production bridge.

For the full architectural map — memory, context survival, attention management — read the series.