My Agent Finally Learned Who Was Worth Calling

Neha Kurudi — Tue, 29 Sep 2026 17:47:36 +0000

For months I noticed that a collections agent could look smart while still making the mistake. Give it an invoice and a customer profile. It could produce a polite reminder. Ask again the day and it would produce another reminder. After several ignored emails it would still suggest sending one more.

That was the behavior I wanted to change.

I built PayRecall, a prototype B2B collections agent that uses Hindsight as its memory layer. The experiment was simple: give the overdue invoice to the same LLM twice. In one run the agent has no customer history. In the other it can remember collection attempts.

The difference was not a prompt. The difference was that the agent finally had a reason to know who was worth calling.

PayRecall starts with an accounts‑receivable problem: an invoice is overdue and someone must decide what to do next. The real decision is not send a reminder." A collector needs to know who to contact, which channel to use what tone is appropriate when the customer is likely to respond what has worked before and what should be avoided because it failed previously.

For the prototype I created collection records containing emails calls, WhatsApp conversations, payment promises, a dispute and a change in the customer’s contact. The main example involved Meridian Retail and overdue invoice INV‑2041 four lakh rupees. The important detail is that the LLM does not receive all this history directly. Hindsight does.

To keep the comparison fair the memory‑OFF path receives basic CRM information such as customer name, industry, city, payment terms and active contacts. It does not receive notes, payment habits or interaction history. As a result the memory‑OFF agent can still make a recommendation. It cannot know what happened previously.

The memory‑ON path retrieves history from Hindsight. Each interaction is stored with tags such as customer, channel, actor, invoice and outcome. Customer tags create a retrieval boundary so that one customer’s history never appears in another customer’s results.

Without memory PayRecall recommends sending an email reminder to the contact. The problem is that history shows this approach repeatedly failed. Previous records reveal that shared accounts inboxes ignored reminders Ramesh Iyer responded well to phone calls and kept payment promises a firm notice triggered a dispute of payment Ramesh left the company and Kavita Menon became the new decision‑maker.

Without memory none of those facts exist from the agent’s perspective.

With Hindsight enabled the agent receives memories, a generated customer playbook and evidence‑based reflections before making a recommendation. Of suggesting another generic email it may recommend calling Kavita during her known availability using a friendly tone.

Importantly the recommendation is supported by evidence. Ignored emails, calls, past disputes, contact changes and communication preferences directly influence the decision. At that point memory becomes more than retrieval. The customer’s history changes the action itself.

Hindsight also generates a customer playbook by consolidating repeated events into higher‑level patterns. The playbook concluded that shared accounts inboxes were ineffective calls worked better than emails firm notices created friction Ramesh was no longer the contact and Kavita was now responsible for payments. This changed memory from a record of events into a tool for decision‑making.

PayRecall also closes the learning loop. After a recommendation the collector records what happened. The system stores the agent’s action and the customer’s response separately. Hindsight extracts facts, updates memory refreshes the customer playbook. Generates the next recommendation.

The cycle becomes: Action → Customer Response → Memory → Updated Playbook → Next Action.

One important lesson came from testing. I recorded a test outcome stating that a customer preferred email. Later the memory‑enabled agent began using that preference. The system had learned what I told it. The mistake was treating a memory write like a test fixture when it actually behaved like application state. As a result PayRecall uses retrieval rules, duplicate checks and cleanup procedures, before demonstrations.

The question I started with was simple: Can an agent remember who is actually worth calling?

Hindsight is not a tool that knows the customer. Hindsight lets the agent collect evidence pull the facts, spot patterns and use those patterns to make better decisions later. The LLM still writes the words. Hindsight keeps the story continuous. With PayRecall Hindsight finally helped decide which customers were worth calling and PayRecall used that insight to focus on the ones.

Hindsight GitHub:
documentation:
Vectorize agent memory:

DEV Community: Neha Kurudi

My Agent Finally Learned Who Was Worth Calling