UniVerse

Lost in the Basement? The AI That Knows Your Campus Better Than You Do.

"University life is complex. Support shouldn't be. We’ve integrated the entire campus experience into one intelligent interface, proving that the smartest thing about a campus should be the way it helps its students."

WHY CONTEXT WINDOW FAILED CONTEXT BOT

"Did it seriously just do that?" I stared at the logs while our agent calmly reminded a student about their 4:00 PM chemistry lab before they could even finish asking about tonight's club flyer. We had just spent the last two days trying to prove that a campus assistant shouldn't feel like a chatbot with a five-second memory.

THE PROBLEM : THE STATELESS STUDENT EXPERIENCE

When we started building the Smart Campus AI Assistant, we thought we could solve everything with a massive context window. The goal was simple: stop students from asking the same repetitive questions about campus events, club meetings, and deadlines.

In the beginning, we tried the "Goldfish Method." Every time a student asked a question, we’d fetch their basic profile, append it to the prompt, and hope the LLM figured it out. It worked for name-dropping, but it failed the moment life got messy. If a student joined five different clubs or had a shifting lab schedule, the context became a noisy soup of conflicting instructions. The agent would "forget" a preference mentioned on Monday when recommending an event on Thursday because that specific token had been pushed out of the immediate attention span.

BEYOND THE CONTEXT WINDOW : ENTER HINDSIGHTS

We realized that for an agent to be "smart," it needs to learn from experience, not just read a static text file. We pivoted from a massive prompt to a memory-driven architecture using Hindsight.

The architecture shifted from "Stateful Prompts" to "Event-Driven Learning." Instead of trying to predict what the student might need, we built the system to retain specific experiences and recall them only when the semantic context demanded it.
How We Implemented Agent Learning
In our core logic, we moved away from generic RAG. We used the Hindsight documentation to implement a pattern where the agent observes user behavior and updates its own "experience" store.

For example, when a student interacts with the bot about a club, we don't just answer the question; we log the experience to help the agent change its strategy based on past interactions:

EXAMPLE OF RETAINING A STUDENTS PREFERENCE

async def handle_club_query(student_id, query_text):
# Logic to answer the specific question
response = await campus_llm.ask(query_text)

# The Learning Moment: Retain the interest for future personalization
if "join" in query_text.lower() or "meeting" in query_text.lower():
    await hindsight.retain(
        user_id=student_id,
        experience=f"Student showed active interest in {extracted_club_name}.",
        metadata={"type": "club_affiliation", "priority": "high"}
    )
return response

By offloading this to the agent memory page on Vectorize, the agent's "brain" stayed lean. It didn't need to hold 20,000 tokens of "maybe relevant" info; it just needed to query for what it had learned about this specific student's interests.

REAL WORLD BEHAVIOUR : THE "AH!" MOMENT

The difference in behavior was immediate once we moved to experience-based learning.

Before Hindsight:
Student: "Where is the IEEE meeting?"
Bot: "Room 302." (The agent repeats the same answer every week without context) .

Student (Two hours later): "I'm hungry, where should I go?"

Bot: "The Dining Hall is open." (Ignoring the fact that the student is currently across campus at the IEEE meeting).

BEFORE HINDSIGHTS :

Student: "Where is the IEEE meeting?"
Bot: "Room 302." (The agent repeats the same answer every week without context) .

Student (Two hours later): "I'm hungry, where should I go?"

Bot: "The Dining Hall is open." (Ignoring the fact that the student is currently across campus at the IEEE meeting).

AFTER HINDSIGHTS :

Student: "Where is the IEEE meeting?"

Bot: "Room 302. I've noted you're attending; I'll silence your non-emergency notifications until it ends."

Student (Two hours later): "I'm hungry, where should I go?"

Bot: "Since you're at the Engineering building for IEEE, the food truck behind the lab is your closest option. I remember you mentioned wanting to avoid the crowded quad earlier".

LESSON LEARNED FRONTHE TRENCHES CONTEXT IS A LIABILITIES NOT ASSETS

The more "junk" you put in a prompt to help the agent remember, the more likely it is to hallucinate or ignore the actual user command.

Memory Requires Filtering: You shouldn't retain everything. We learned to only store "High-Impact Experiences"—things that change how the agent should act in the future, like club memberships or recurring deadlines.

The "Skeptical Developer" Test: We initially doubted if a third-party memory layer was necessary. We tried building our own vector store first, but the "learning" part—ranking past experiences by relevance to a current, ambiguous query—is harder than it looks. Hindsight handled the heavy lifting of semantic recall so we could focus on the campus logic.

PROJECT GALLERY 📷📸