When building stateful AI applications, developers usually default to a traditional relational database or vector store combined with heavy prompt engineering to keep track of user preferences. While building BrandPulse—an intelligent marketing platform for programmatic digital out-of-home (DOOH) campaigns—I ran into the limits of this approach. Brands constantly update their target audiences, campaign goals, tone of voice, and safety constraints. Storing these guidelines in static tables meant our agents constantly suffered from context drift or required massive, error-prone prompt stuffing on every API call.
To solve this, I ripped out our traditional session storage and integrated Hindsight directly into our FastAPI backend. In this article, I will break down why treating memory as an isolated, streaming agent bank changes the way we handle state, and how it simplifies dynamic context management.
The Problem With Traditional State Management
In standard architectures, updating a user's brand profile means executing an SQL UPDATE statement, and then manually pulling and formatting those rows into system prompts whenever an LLM call is made. As applications scale and business rules evolve, this manual plumbing becomes brittle:
Context Bleed: Multiple clients or campaigns easily contaminate each other's parameters if namespaces aren't strictly isolated.
Rigid Schemas: Storing unstructured preferences (like nuanced brand tones or complex safety filters) into fixed columns limits the agent's ability to reason over them semantically.
Prompt Bloat: Passing entire rulebooks on every request inflates token usage and latency.
By exploring agent memory via Vectorize, I realized that agent state shouldn't be treated like static CRUD data. It needs to be continuously retained, recalled, and reflected upon dynamically.
System Architecture and Flow
BrandPulse acts as an automated strategist for programmatic advertising campaigns. Brands log in with their core parameters—such as industry, target audience, campaign goals, tone of voice, blocked topics, and geographic target locales.
Instead of writing these parameters to a rigid relational database where they remain static, the system routes each profile update into an isolated Hindsight memory bank keyed by the brand's unique identifier. When a user submits or updates their profile, the backend performs three core operations:
Retain: Ingests unstructured and structured brand constraints into the Hindsight memory stream.
Recall: Pulls relevant context dynamically based on incoming strategic queries.
Reflect: Synthesizes high-level marketing suggestions and specific programmatic screen placements tailored to the brand's precise profile.
The system is split into a lightweight asynchronous FastAPI backend handling API requests and a responsive single-page frontend interface communicating over REST endpoints. According to the Hindsight documentation, memory banks maintain semantic indexing over time, allowing agents to reason over historical preferences rather than just matching keywords.
Code-Backed Implementation
Here is how clean the ingestion and reflection flow looks in our FastAPI backend using the official Python client:
import os
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from hindsight_client import Hindsight
app = FastAPI(title="BrandPulse AI API")
app.add_middleware(
CORSMiddleware,
allow_origins=[""],
allow_credentials=True,
allow_methods=[""],
allow_headers=["*"],
)
HINDSIGHT_BASE_URL = os.environ.get("HINDSIGHT_BASE_URL", "https://api.hindsight.vectorize.io")
HINDSIGHT_API_KEY = os.environ.get("HINDSIGHT_API_KEY", "")
client = Hindsight(base_url=HINDSIGHT_BASE_URL, api_key=HINDSIGHT_API_KEY)
class BrandProfile(BaseModel):
brand_name: str
industry: str
target_audience: str
campaign_goals: str
brand_tone: str
keywords: str
blocked_topics: str
target_locales: str
@app.post("/api/brand/login")
def save_brand_profile(profile: BrandProfile):
bank_id = f"brand-{profile.brand_name.strip().lower().replace(' ', '-')}"
profile_content = (
f"Brand Name: {profile.brand_name}. Industry: {profile.industry}. "
f"Target Audience: {profile.target_audience}. Campaign Goals: {profile.campaign_goals}. "
f"Brand Tone: {profile.brand_tone}. Keywords: {profile.keywords}. "
f"Blocked Safety Topics: {profile.blocked_topics}. Target Locales: {profile.target_locales}."
)
# Retain the profile inside the brand's isolated memory bank
client.retain(bank_id=bank_id, content=profile_content, context="brand-profile-login")
# Reflect to synthesize strategic marketing placement recommendations
reflect_res = client.reflect(
bank_id=bank_id,
query=f"Synthesize strategic marketing suggestions and programmatic digital screen placements in {profile.target_locales} matching the tone '{profile.brand_tone}' and goals '{profile.campaign_goals}'."
)
return {
"status": "success",
"bank_id": bank_id,
"strategic_recommendation": getattr(reflect_res, 'text', str(reflect_res))
}
Before / After Example
Before (Static Database + Manual Prompt Stuffing):
Profile updates required writing custom SQL migrations for new fields (e.g., adding blocked_topics).
System prompts grew to over 2,000 tokens just to inject static rulebooks, leading to high latency and missed contextual nuances.
After (Hindsight Memory Stream):
Unstructured profile data is retained instantly into an isolated memory bank without changing table schemas.
Reflection endpoints dynamically synthesize strategic recommendations based on historical memory, reducing prompt overhead and improving output relevance.
Results and Behavior in Practice
When testing this implementation with luxury or tech brand profiles, the difference in agent output is immediate. Rather than spitting out generic placeholder text, the reflection call queries the retained memory stream to produce highly tailored outputs—such as recommending high-impact programmatic digital out-of-home (DOOH) screen placements in premium corporate tech parks and luxury malls during peak evening commutes, while strictly adhering to safety filters that block controversial topics.
Dead Ends and Lessons Learned
Decouple State from Schemas: Letting an external memory engine handle unstructured profile data saves countless hours of database migration work.
Deterministic Bank IDs: Using clean, predictable identifiers like brand-{name} ensures secure, isolated memory scoping per client.
Reflect Over Raw Retrieval: While pulling raw history is useful, utilizing reflection endpoints lets agents synthesize actual business logic directly from stored context.
Resilient Error Handling (A Dead End We Hit): Early on, unhandled network hiccups during external memory calls caused silent failures that defaulted to mock data. Implementing robust exception handling and fallback logging solved this immediately.
vectorize-io
/
hindsight
Hindsight: Agent Memory That Learns
What is Hindsight?
Hindsight™ is an agent memory system built to create smarter agents that learn over time. Most agent memory systems focus on recalling conversation history. Hindsight is focused on making agents that learn, not just remember.
hindsight-learning-demo.mp4
It eliminates the shortcomings of alternative techniques such as RAG and knowledge graph and delivers state-of-the-art performance on long term memory tasks.
Contents
- Memory Performance & Accuracy
- Quick Start — server · clients · platforms · embedded
- Adding Hindsight to Your Agent — LLM Wrapper · integrations · coding agents · MCP
- Core Concepts — memory types · retain / recall / reflect · observations · mental models & knowledge pages · banks
- Use Cases
- Running in Production
- Resources
Memory Performance & Accuracy
Hindsight is the most accurate agent memory system ever tested according to benchmark performance. It has…
Overview | Hindsight
Why Hindsight?
class="c-embed__favicon m-0 mr-2 radius-0"
src="https://hindsight.vectorize.io/img/favicon.png"
loading="lazy" />
hindsight.vectorize.io
What Is Agent Memory? A Complete Guide | Vectorize
Agent memory lets AI agents retain, recall, and reflect on experience across sessions. Learn how it works, the key memory types, and how to implement it.
class="c-embed__favicon m-0 mr-2 radius-0"
src="https://vectorize.io/icon.png?cb267cb4d8dc986d"
loading="lazy" />
vectorize.io
){kind=link}
Top comments (0)