Memory & Agency - Build the First LLM Agents That Remember

25 Sept 2025 • Facilitator: Performics Labs Newsroom

💡 TL;DR - We’re moving from static SEO rules to living agents that remember, reason, and grow. Instead of building brittle optimizers, we’re designing assistants that learn over time - balancing fast, reflexive insights (via Chrome’s built-in AI) with deep, long-term reasoning (via AWS cloud agents). This is your chance to join two parallel hackathons and help prototype the first generation of memory-enabled LLM agents.

🌟 The Paradigm Shift: From Rules to Remembering

Traditional SEO/Optimization: Rules → Application → Short-term Gains → Obsolescence
Memory-Based Optimization: Context → Experiment → Memory → Adaptive Growth

We’re moving from “apply fixed best practices” to “build systems that remember what worked, why it worked, and when it stopped working.” The simple idea: agents that justify their knowledge can adapt faster than those that only recall patterns.

The Two Siblings

⚡ Chrome Agent (System 1) - fast, reflexive, on-device. Suggests optimizations instantly while you edit content.
☁️ AWS Agent (System 2) - slow, deliberative, in the cloud. Consolidates experiences, runs causal tests, and updates strategies over time.

Together, they create an epistemic loop: fast reflexes informed by long-term wisdom.

🗺️ How to Navigate This Page

Watch the 7‑min explainer video (see top of page).
Pre-read the anchor article: Part 2 - Memory & Agency.
Review context articles:
- Part 1 - The Phenomenology of Search
- Part 3 (coming soon) - User Behavior & Personas
Discuss & vote in comments at the bottom of this page.
Join a hackathon track - see below.

🛠️ Build-In-Public Sprint

Goal: Ship a minimal, testable memory-enabled optimization agent that demonstrates the epistemic loop: fast Chrome reflexes + slow AWS reflection.

Track A - Chrome Reflexes (System 1)

Step	What to Deliver	Hints
1	Chrome Extension Starter	Use Chrome AI APIs.
2	Working Memory Buffer	Maintain 5–9 live insights (Miller’s Law).
3	Pattern Cache	Cached successful optimizations, updated on-device.
4	Fast Suggestions	<2s content rewrites and recommendations.
5	Feedback Logging	Capture user accepts/rejects for cloud sync.
6	Demo	90-sec screencast of real-time suggestions; repo + README

Track B - AWS Memory Palace (System 2)

Step	What to Deliver	Hints
1	Cloud Memory Service	Episodic + semantic + procedural memory tiers.
2	Reinforcement Learning Core	Multi-armed bandits with contextual features.
3	Causal Inference Layer	Separate correlation from causation.
4	Heuristic Update Loop	Send refined strategies back to Chrome.
5	Conflict Resolution	Partition truths by context (e.g., B2B vs consumer).
6	Demo	90-sec video + repo + README showing adaptive learning over time.

Recognition: Submit to one or ideally both hackathons for external prizes and showcase internally as pioneering epistemic-agent innovation.

🏁 Devpost Links

Chrome AI Hackathon: https://chromeai.devpost.com/
AWS Agents Hackathon: https://aws-agent-hackathon.devpost.com/

Tip: Frame your submission around memory, justification, and adaptation, not just surface‑level metrics.

✅ Acceptance Criteria & Scoring

Required

Epistemic loop: Demonstrate fast suggestions + slow consolidation.
Working memory limits: Respect cognitive science constraints (7±2).
Justification chains: Store provenance, reliability, and defeaters.
Adaptive learning: Show strategy updates over time.

Bonus

Transparency: Visualize uncertainty and memory reasoning.
Cross-domain transfer: Apply strategies across contexts.
Resilience to interference: Handle conflicting evidence without collapse.
Open-source commitment: Make your repo usable for future builders.

🤔 Discussion Prompts

What does it mean for an AI agent to “remember responsibly”?
Should we prioritize preservation (facts with provenance) or generation (patterns from fragments)?
How do we design exploration vs. exploitation in marketing contexts?
What are the risks of over‑trusting memory, and how do we mitigate them?
Could a network of agents share memories without collapsing into noise?

📚 Pre-reads & Reference Pack

Epistemology & Memory

Information Theory & Learning

Agent Design

🚀 Ready to Build the First Remembering Agents? This is your chance to prototype assistants that don’t just optimize - they remember, justify, and adapt. Join the hackathons, test ideas in public, and help launch a new era of epistemically grounded AI.