Skip to content

Sponsors

Six sponsor tools carry the build. Each section says what the tool does in Memento Emptor, then what we ran into while building on it. The feedback comes from the research records the build was planned on, linked under each one.

What it does here: the evidence, and the action.

  • Webhooks for captures, refunds and disputes reach the layer with no agent involved. Every delivery is verified with PayPal before anything is stored.
  • Dispute reason codes become classified reasons. A refund’s reason comes from the shopper, classified by Jev.
  • Evidence is read under the merchant’s own token, which is why a merchant runs the layer beside its PayPal app.
  • The Agent Toolkit’s orders tools make the approved sandbox purchase, with explicit shopper approval and idempotent request ids. The layer itself never moves money.

Feedback.

  • The webhook event simulator (POST /v1/notifications/simulate-events) answered 404 with an empty body on both sandbox hosts, so we drove real events instead.
  • No API available to our app showed a sandbox merchant’s home currency. A “seller doesn’t accept payments in your currency” refusal cost us a replacement sandbox account.
  • Manual resend delivered within about a minute. We did not observe an automatic retry in the sandbox.
  • A refund carries no reason. That gap is why this layer exists, and a reason field on refunds would help every agent built on the toolkit.

Research: dev-diary/research/paypal.md.

What it does here: both data screens.

  • The shopper’s evidence ledger: beliefs, the evidence behind each, status labels, the seeded mark, and contest.
  • The merchant app: returns per product line and size, by reason and direction.
  • AG Grid Community 36.3.0, themed for light and dark with Chaaya tokens.

Feedback.

  • The bundle is about 1.2 MB minified (340 KB gzip), so the ledger loads the grid lazily rather than on the landing page.
  • Version 36 themes apply through adoptedStyleSheets, which jsdom does not implement, so theme tests stay structural.
  • Tests must wait for firstDataRendered, and a container measured at zero width renders only the first column until column virtualisation is turned off.

Research: dev-diary/research/ag-grid.md.

What it does here: real products for the agent.

  • The reference agent searches for candidate products and variants from real retailers, with buy links.
  • Per-SKU fit labels (top-fit) tell the layer when a new shirt shares a cut with the ones the shopper returned. The size comes from recall, not a size chart.

Feedback.

  • The fit vocabulary is per category: top-fit for shirts, footwear-cut and shoe-collar-height for shoes.
  • Two SKUs in one product line can carry different fit labels, so the line name alone does not prove a shared cut. Comparison has to be per SKU. A normalised cut field across categories would help.

Research: dev-diary/research/channel3.md.

What it does here: the pool.

  • The pooled fit index, fit-rows, holds one structured row per settled purchase. A row carries a salted per-line shopper hash and never a shopper’s identity, free text or order detail.
  • Aggregations by product line, size, reason category and direction, with exact distinct-shopper counts (cardinality), decide whether a line is Building or Earned. This is what gets a new shopper right on day one.
  • The merchant app reads the same index through product_fit, never directly.
  • Elastic’s jina-embeddings-v3 inference endpoint also embeds Lambo’s memory.

Feedback.

  • Serverless refuses external versioning on the partial _update API, so each row is a whole-document write with version_type=external_gte, versioned by the purchase’s state.
  • A cold first connection from our services took about 6.9 s, against about 236 ms warm, so recall keeps a warm connection and a 20 s timeout.
  • The project is a Vector Database project with an active licence, not a trial. The trial-status endpoint answers 410 in serverless mode, which did not make that obvious.
  • Serverless runs vector indices with source vectors excluded, which changed how Lambo could re-score results.

Research: dev-diary/research/elastic-fit.md and dev-diary/research/elastic.md.

What it does here: the agents, and the proof they behaved.

  • Both shopping agents run as Astropods deployments: the reference agent, and a second agent with its own prompt that never saw the first conversation.
  • Model calls go through the platform gateway, so no personal model key is needed. The agent key and the Channel3 key arrive from the platform secret store by reference.
  • Every turn’s trace shows its tool spans. That is the independent proof an agent called recall_for_purchase before recommending, and honoured a correction it did not see made.
  • Evaluators score traces, a human review is saved per trace, and traces of interest become dataset items.

Feedback.

  • ast agent restart --component demanded a component and then rejected both the component name and the full workload name, so we redeployed instead.
  • A deployed spec must declare agent.interfaces.messaging. Without it the chat host had no web sidecar and returned 404 behind a sign-in redirect that looked healthy to anonymous checks.
  • A no-login page needs the console’s Public switch. The CLI’s access options offer signed-in users only.
  • Scripted calls through the messaging proxy need a browser user agent to pass the edge’s bot check.
  • The free tier’s signup credit covered the whole build.

Research: dev-diary/research/astropods.md.

What it does here: hosting.

  • One Docker web service runs the edge, the layer, the ledger app and the Lambo memory hub on a persistent disk.
  • The merchant app runs as a second service.
  • Render hosts the public PayPal webhook endpoint, and a push to main deploys automatically. Keys live in environment variables, never in the repository.

Feedback.

  • An environment variable change does not redeploy on its own. A new key sat behind the old instance until we triggered a manual deploy.
  • A disk-backed service runs one instance and loses zero-downtime deploys, so each deploy briefly drops webhook deliveries. PayPal’s resend and event-id idempotency cover the gap.

Research: dev-diary/research/hosting.md.

  • Lambo: graph memory, one session per shopper.
  • Jev via OpenRouter: the reason classifier.