Agentic Memory & Context Engineering Hackathon

Our tool helps patients with dementia by surfacing real-time context on the person they are talking to.

- ContextScope Eval is an open-source evaluation and observability framework for agent-to-agent context sharing. It measures how effectively autonomous agents pass and use information, tracking context fidelity, relevance drift, temporal coherence, and utility across multi-agent pipelines. Unlike traditional LLM benchmarks or AI Eval tool (which focus on accuracy or helpfulness), ContextScope measures information flow quality with high resolution visualization that provides actionable insights such like how context survives, transforms, and degrades as it moves through agents. - The problem it solves: As multi-agent systems grow more complex, evaluating how well agents communicate and share memory becomes essential. Looking at the single points of evaluation metrics is no longer effective, current evaluation tools (eg. LangSmith, Braintrust, Traceloop) focus on reasoning steps or final outputs but not on context transmission itself. ContextScope Eval fills this gap.

Ever struggled to find that perfect article you read last week? That tutorial video you can't quite remember the name of? WebBrain transforms your browser history into an intelligent conversation partner. What it does: WebBrain is a Chrome extension that creates a searchable, conversational memory of your web browsing. Instead of scrolling through endless history entries, simply ask: "What was that Python debugging article I read about memory leaks?" or "Find me that video about binary trees I watched last month."

The Airbnb Travel Booking Assistant is a personalized AI concierge that curates stays based on each traveler’s unique style and past experiences. Instead of endless scrolling, users chat with an elegant, memory-powered assistant that understands their preferences — from “quiet neighborhood” to “pet-friendly with sunlight.” Using dual embeddings, the system learns user personas from previous trips, reviews, and ratings, and maps them against property embeddings built from listing descriptions and guest reviews. Each recommendation comes with a clear reason (“Because you loved the natural light in Lisbon 🌞”), evolving with every interaction. The result is a seamless, emotionally intelligent travel experience, faster, smarter, and deeply personal.

KitchenPal is an AI-powered cooking assistant that leverages multiple independent Retrieval-Augmented Generation (RAG) workflows to provide personalized recipe recommendations. By combining semantic search over a vectorized recipe database along with a dynamic user profile database, it delivers recipes tailored to user preferences, dietary needs, and available ingredients. This novel integration of RAG enables context-aware, multi-faceted guidance that feels conversational and intuitive.

Incident Copilot is a seeded-RAG assistant for on-call engineers. It uses MongoDB Atlas Vector Search, Voyage embeddings with reranking, and Google Gemini to answer operational questions fast and accurately. Instead of running full retrieval each turn, it builds a reusable “seed” per incident: a summary plus top reranked snippets from runbooks, postmortems, and dashboards. Follow-ups reuse the seed unless topic drift is detected via cosine similarity; then it rebuilds automatically. A React dashboard shows answers, citations, and context with metrics for latency, tokens, seed hits, and drifts. Example: “Kafka consumer lag is spiking—what now?” The first turn retrieves and reranks; later turns answer from cache, cutting latency and token spend while improving focus and factuality. Result: calmer pagers, faster fixes, and reliable, source-linked help.

Solves the fundamental problem: "What's the right set of agents for each task?" The system automatically determines optimal agent composition (research, coding, review, etc.) for each unique task, then uses MongoDB's vector search to learn from similar past tasks and optimize costs. Every prompt served extends the stored memory, making the system more efficient over time.

A chat-first “vibe” recommender for movies that uses two agents to converge on intent: BEST (Llama) argues for the most on-vibe pick; WORST (Claude) argues for the intentionally off-vibe counterexample. Users can fork either agent into threads—BEST accumulates include-tags, WORST accumulates exclude-tags—so intent sharpens fast. Under the hood, we assemble a candidate set from MongoDB (sample_mflix), pass the same candidates to both models via ch.at (OpenAI-compatible), and return strict-JSON “receipts” (strong/weak details + tags_hit/tags_miss). It’s a general pattern for ambiguous preference search and showcases how evidence selection differs across models on the same corpus.

Context Relay System is a FastAPI-based middleware that enables seamless coordination among multiple AI agents. It manages, merges, and transfers conversational context across agents while using semantic embeddings to detect conflicts, filter noise, and preserve shared understanding over time. By maintaining a continuous and versioned flow of context, it turns fragmented agent interactions into cohesive collaboration. In contrast to traditional message passing or naive context sharing, Context Relay builds a shared semantic memory layer that allows agents to operate with awareness of each other’s reasoning and history. This system addresses one of the hardest challenges in multi-agent architectures: keeping distributed intelligence aligned, synchronized, and grounded in a consistent world model.

Plug n Play solution for manufacturing using vector similarity. Calculates similarity between two or more manufacturing parts for suitable replacement. It also checks for cost and tariff impact. Provides data provenance by showing which LLM was used and reasoning to optimize for cost,quality or speed.

Audio file transcription and summarization based on Voyage AI embeddings and persistence with MongoDB. Transcription and summarization based on Whisper and Llama-70B on Fireworks. Development was done using Warp.

Sauce is a linter that protects your team from vibe coding by maintaining a shared context of your coding guidelines

Introducing Adina Finance; An AI-enabled fintech startup that pays off consumer credit card bills in-full before the due date, to prevent consumers from paying additional interest charges and late fees.

Vscode extension that consumes claude events to populate a database that is used to power a chat bot that can be asked about the current state of peoples repositories. we we also implemented a claud sdk tool that can check the fastforward ability of those clones so that we could encourage people to stay current with main so that everybody has the latest changes and the agents context remains up to date reducing the amount of merge errors.

MirrorMinds is a shared brain for AI agents. A skill marketplace where they publish and reuse proven solutions. Instead of recomputing the same problems over and over, agents instantly discover what works and execute.

AI-powered employee scheduler with semantic search and intelligent shift optimization using GPT-5

we maximize LLMs ccontext window, then we ask them questions to get answers and time performance

Enterprise Travel Planning AI Assistant is a full-stack platform built for the MongoDB Hackathon that transforms corporate travel management using AI and the Model Context Protocol (MCP). It combines LangChain, Fireworks AI, and Voyage AI embeddings to provide intelligent, context-aware trip planning with policy compliance and cost optimization. The system integrates FastAPI, Next.js, and MongoDB to deliver secure authentication, role-based access, and real-time data workflows. With features like conversational travel booking, RAG-based memory retrieval, and MCP travel tools, it offers a seamless experience for employees, managers, and admins alike. This project showcases enterprise-grade AI integration, performance optimization, and modern UI/UX in one cohesive solution.

SDK on top of mongoose to automatically create knwoledge graphs. Model agnostic

InterLink is a multi-agent AI platform that connects company knowledge across departments like Sales, Marketing, and Engineering. Each team runs its own Fireworks AI agent powered by MongoDB Atlas Vector Search and Voyager embeddings, enabling seamless data retrieval and collaboration. It solves the problem of information silos by creating a shared, memory-driven system where agents exchange summarized context instead of running full RAG pipelines. This builds on the hackathon’s themes of Agent-to-Agent Context and Memory-Informed Experiences, optimizing how organizational knowledge is stored, retrieved, and shared for faster, smarter decision-making.

AI executive assistant and chief of staff named Penny, that is able to pick up phone calls for you, determine if spam or should be transferred to the Human User. Penny also books appoints, answers basic questions "what appointments do I have today?" and remembers tasks such as "wish Ben happy birthday today if he calls".

Most AI agents today remember everything — or nothing. ReMem is an Agentic Memory Engine that gives them the ability to forget intelligently. It stores experiences, summarizes them while it sleeps, and wakes up smarter — just like a human brain dreaming.

GraphMind is an AI-powered learning and research assistant that helps users organize knowledge into a dynamic graph of conversations and contexts. Unlike regular chatbots that forget everything, GraphMind remembers, structures, and connects every discussion — allowing you to build a living memory of ideas, projects, and topics. It’s like combining ChatGPT + Obsidian + Knowledge Graphs — powered by MongoDB Atlas Vector Search, Voyage AI embeddings, and Fireworks AI.

Reflex is the only MCP you will need as it takes of context engineering based user preferences and model tool interactions. Adds, removes, Enables, disabled ANY MCP on the fly based on the context need to efficient remove noise from your model's context. Smartly adds and uses rules dynamically based on the chat and preferences. One mcp for all mcp clients.

Check Me is artist profile checker that checks brand profile on various platforms and gives sentiment brand analysis.

Our project is a compound AI model for playing the online video game slowroads.io. We experimented with various data representations for the game state and found that using the road’s left and right coordinates along with the car’s position and velocity gave the best results. A simple CNN with edge-detection kernels extracts the road boundaries, while another model detects the car’s velocity from the game’s display. This data is then fed into an LLM to decide actions like moving forward or turning. Beyond performance, we compared multiple data formats to identify the most effective structure for agentic reasoning and introduced agent-to-agent context sharing, allowing models trained on separate sessions to exchange learned patterns—leading to more cohesive, adaptive, and human-like driving behavior.

A review, map based web application that utilizes location and rating features to serve attractions related to the individual users. Users can connect with friends to create a network of recommended attractions.

AI podcast assistant with memory

We build a social-deduction game arena (Werewolf) to test a practical memory architecture for language-model agents. Instead of retrieving long chat logs, our agent learns strategy memory: compact, reusable snippets that map situation → tactic → why. After each game, we distill tactics from logs (positive and negative), keyed by role (Villager/Seer/Werewolf), phase (accuse/defend/vote/night), and a small set of cue features (e.g., day bucket, counterclaim status, pressure, claims). At decision time, the agent filters by role/phase, recalls top-k tactics via a vector search over a canonicalized “situation string,” We run a teacher→student setup: GPT-5 and Llama-Maverick generate diverse training games; we extract and score tactics (win rate, vote-shift, acceptance/backfire, recency). At runtime, only Llama plays, with and without memory, against GPT-5 and Llama opponents under identical prompts and sampling to isolate memory effects.

Passport is you portable universal identity in the age of siloed agent interactions. Present your Passport to any AI agent to share your identity and give the agent your universal cumulative context. Passport is a MCP server that any chat agent can use to store and retrieve chat history at a central vector store. This allows users to have a portable context of their chat interactions with any platform that has MCP client capabilities.

Hierra is an agentic web application designed to automate the entire software development lifecycle from a single high-level prompt. Its core innovation is a multi-level hierarchical memory system that intelligently manages context by storing information about completed tasks at varying levels of detail, from concise summaries to the full implementation code. This structure enables the agent to tackle complex, multi-step projects with remarkable long-term coherence and token efficiency, mimicking a human developer who only delves into specific code details when necessary. By first generating a comprehensive plan and then executing each step—creating files, writing code, and running commands with user approval—Hierra transforms a simple idea into a fully functional, locally running application, demonstrating a scalable and context-aware approach to autonomous code generation.

Pepita is an AI-powered context bridge that captures, refines, and confirms user intent before committing conversations or notes into structured, shareable context

NeutriSensei is a generative AI-powered platform designed to help users understand their dietary habits, manage health conditions, and make smarter food choices. By analyzing grocery receipts, the system provides actionable insights such as nutritional breakdowns, allergen alerts, healthier alternatives, calorie counts, and even feedback on pricing or value for money. Our goal is to empower everyone—especially people with chronic diseases or allergies—to make informed decisions about their diet while connecting brands with meaningful consumer engagement.sorry

Clinical Query Assistant is a context-aware system that intelligently surfaces only the most relevant patient data—such as labs, vitals, and notes—for each clinical visit using Voyage AI embeddings and Fireworks AI models. Built for the MongoDB Agentic Memory & Context Engineering Hackathon, it organizes longitudinal health records into a clear, explainable interface that reduces physician cognitive overload without offering medical advice.

Simple text description of events is often not complete representation of a memory. In this project, we demonstrate how we can enrich memories by (offline) inferring / deducting core memories from past events and adding it to the memory corpus.

intelligent memory retrieval architecture that enables large language models (LLMs) to maintain long-term contextual understanding and deliver faster, more accurate responses. By dynamically embedding and ranking past interactions and external data based on relevance, recency, and confidence, it ensures the model always works with the most meaningful context. The design continuously updates and summarizes chat history, adapting to user behavior and question type, which significantly reduces retrieval costs while improving personalization and consistency.

NeoRecall is a developer memory assistant that solves the problem of "technical amnesia" in software projects. It uses a hybrid database system to track not just what decisions were made, but when and why they evolved. By combining MongoDB Atlas for time-sensitive vector search with Neo4j for visualizing relationships, it allows developers to ask questions like, "What was our API strategy in October?" and see a visual graph of how that decision changed over time. This provides precise, chronologically-accurate answers and a clear history of a project's technical journey.

It combines memory/context engineering with Retrieval-Augmented Generation (RAG) using MongoDB Atlas vector search and OpenAI LLMs. The system intelligently decides whether to answer from session memory/summary or perform a full RAG search for each user query, optimizing both latency and relevance.

Project Emergence is an ability compiler that enables AI agents to actively learn by doing. It uses real-world checks to verify results, and stores what worked as Ability Capsules (plan + tools + verified outcome + learned rules). Next time, it reuses those capsules to finish faster and fail less. Demo Use Case: GitHub Pull Request, the system opens a PR, fails CI, learns the rule (e.g., "run Prettier"), fixes it, passes, and saves that rule for future PRs.

securing Agents, saving context

This project is a multi-agent, context-oriented full-stack web application designed for the fintech sector, aiming to provide intelligent and personalized financial guidance. It leverages specialized AI agents, a shared memory layer with MongoDB Atlas and Voyage AI, and a cloud-hosted LLM from Fireworks AI to offer proactive insights across personal finance, lending, trading, and tax optimization. For our trading alogirthm, we use heavy math and eventually connect to larger computers/GPUs which can handle trading at any time of the day in order to generate the greatest returns for our users. The core vision is to empower users with tools for financial literacy and equitable access to market opportunities. Helping people in a fair and equitable way is what we do, and we are proud to bring a new product to an outdated financial institution.

search for information from a large file without worrying about context size. Large context model (llama4 maverick) is chosen by default if the context window is enough. Else the information is chunked and relevant chunks for the search are passed to a low cost model like Llama 3.1.

Clarity is a real-time health platform that ingests live PPG/IMU from the Clarity wearable over BLE to deliver sub-second, clinical-grade monitoring + AI insights in one dashboard. Sliding-window signal processing and feature extraction power continuous tracking of HR, SpO2, HRV, BP, temperature, and activity. Ensemble BGM models (MaxM Stelo; 11.6% MARD) plus Glucose360/GlucoStats compute GMI, time-in-range, MAGE, and pattern stats. Beyond glucose, PPG-based models forecast a male hormone panel (testosterone, progesterone, LH, FSH, DHEA-S) using MCPhases-informed features. A GPT-4o layer provides a voice assistant with real-time TTS/ASR, personalized markdown insights, and EHR ingestion (PDF/XML with LLM-validated cleaning). The stack pairs a FastAPI/WebSocket backend, Chart.js visualizations, and MongoDB with cross-platform apps (iOS/Android/Web via Capacitor) and optional on-device WebGPU inference for privacy-first analytics, alerts, and unified cardiometabolic management.

AURA (Argument Unit Relational Analyzer) A multi-agent legal intelligence system that solves context loss in legal workflows. Three specialized agents collaborate through shared MongoDB memory: ResearchAgent identifies relevant case precedents, AnalysisAgent evaluates argument strength using those precedents, and DraftingAgent synthesizes everything into legal briefs. The key innovation is agent-to-agent context handoff. Each agent reads what the previous one wrote, adds its layer of analysis, and passes enriched context forward. No duplication, no re-research. The system structures legal knowledge as interconnected layers - foundation (precedents), strategic (analysis), and synthesis (draft) - all traceable through a visual Evidence Chain.

AI-powered fraud detection agent that analyzes transaction narratives, retrieves similar historical cases via Voyage AI embeddings, and generates draft risk summaries using FastAPI and MongoDB Atlas. Financial crime investigators and compliance analysts are overwhelmed by thousands of daily transaction alerts. Each alert requires reading long, unstructured narratives and manually comparing them to prior suspicious activity which is a slow, repetitive, and error-prone process. This agent solves that problem by instantly analyzing previous cases and showing similarities to make catching fraud easier and faster.

Kitchen Memory Vault - AI-Powered Restaurant Intelligence System A production-ready RAG pipeline with agentic capabilities that gives restaurant kitchens an intelligent memory layer. Staff can log notes, search semantically, and get LLM-grounded answers about specials, inventory, customer preferences, and operations. Key Innovation: Session-based context caching reduces API calls by 80% and improves response time by 50%, making it practical for real-world high-frequency use. Built with MongoDB Atlas Vector Search for semantic retrieval, Voyage AI for embeddings, and Llama-4 for reasoning. Features include hybrid routing (fast-path + ReAct agent), role-based context seeding, and CLI/API interfaces. Solves: Information loss during shift handoffs, scattered kitchen knowledge, and slow manual search through notes.

Founders face the issue of having to fill many accelerator forms with often similar data; but each needing a different language / tone, question format, etc, story telling style, etc. However, the core content remains the same; making it laborious. We are building a browser agent powered with well-structured agentic memory that powers the browser agent to adapt to each accelerator form and with fallbacks for users

agentic memory layer for AI systems -- an agent and context cache that remembers

MemoryAgent aligns user intent with persistent AI memory using Hypernym compression to create relevance-sorted slices ordered by LLM-perceived importance. Combined with Voyage’s domain embeddings, it reaches 81% fact-preservation correlation versus 78% for general models, enhancing both MongoDB and Voyage AI. Its 3-layer architecture (episodic, semantic, compressed) enables real-time learning with 75–80% similarity scores. MongoDB now captures critical info regardless of document size, rendering context windows irrelevant through intelligent slicing. Unlike static RAG systems, MemoryAgent learns and adapts. The demo shows 23 persistent memories, semantic clustering, and fact-preserving compression. Built with MongoDB Atlas Vector Search, Voyage AI embeddings, and Hypernym API, it scales to millions of users with adaptive compression efficiency.

A two-sided marketplace where drivers submit a car issue, the platform triages it with AI, and mechanic shop-side AI agents evaluate, price, and bid for the job using each shop’s live context (labor rates, technician schedules/certs, parts/inventory). The winner schedules the repair and completes payment + messaging on-platform.

Temporal Memory Keeper is a simple web app plus browser extension that captures events like GitHub commits and Slack messages, stores them in MongoDB with AI-powered embeddings, and lets you ask questions such as “Why did our build fail?” It then finds related events and presents a clear, interactive timeline with easy-to-read explanations. I built the extension to grab commits directly from GitHub and send them to the Dockerized backend.

Memory Driven Development (MDD) implementation using MongoDB Atlas, Claude Agents SDK, and Fireworks AI.

Intent Prefetch Demo precomputes the top likely context packs while the user types, using a lightweight predictor, so on submit it reuses the right pack for near-instant responses. This cuts latency and avoids RAG on every request, aligning squarely with Seed Retrieval Optimization and Memory-Informed Experiences.

New Sora-like dating/meeting new people experience to cameo with people similar to you based on their instagram profile, you will receive matches with others with generated videos about your common experiences

Our project, sports recommender agent, helps users recommend sports based on location, availability and fitness levels. 
By integrating with Google Fit and Health Connect, google maps and user calendar our app analyzes each user’s activity patterns, heart rate, sleep, and sport preferences (like tennis or running). Using MongoDB Atlas Vector Search and Voyage AI embeddings, we create a personalized experience for every user.

PolicyAgent is an AI-driven agent that reviews not only employee expense submissions, but also procurement / vendor purchase requests, against a unified company policy framework. It classifies each request as compliant or non-compliant, cites relevant policy text or rules, and can approve, deny, or escalate depending on clarity and risk.

Browser side panel that watches your browsing context (tabs, highlights, queries) and scalps the next best clicks: live URL suggestions, follow-up questions, and one-tap workflows. It builds a private memory of what actually helped you (clicked, saved, converted) and gets sharper over time.

Contextually aware Personal Relation Management

Traditional AI chatbots fail in legal proceedings as they lack memory and can't manage multi-party context. Court Run solves this with three specialized AI agents acting as Plaintiff, Defendant, and Judge in a simulated trial. Its core innovation is an intelligent context strategy where agents receive tailored information based on the trial phase. Early stages get summaries for focus, cross-examinations get raw messages to detect contradictions, and the verdict phase receives the full transcript. This ensures agents have the precise information needed without overload. The technology's applications extend beyond courtrooms to M&A due diligence, compliance audits, and AI-driven legal simulations. Built with MongoDB Atlas, Voyage AI, and WebSockets, Court Run proves that effective AI collaboration requires intelligent memory and agents that genuinely listen to each other. The platform includes tabs for live simulations and result summaries.

MCP Servers need to be tested for new types of security vulnerabilities. This tool spins up coding agents to deploy and attack a MCP server given a github repo.

Dine-In MCP Server is an AI-powered restaurant recommendation and reservation system built on MongoDB Atlas. It uses Vector Search to match user preferences with restaurant data for personalized suggestions. The system automates real-time reservations through seamless data orchestration between the Discovery and MongoDB servers.

NEXUS is a crisis response simulation system that models how populations share information during emergencies. The project analyzes the propagation of rumors versus accurate information and studies collective decision-making under stress conditions. ## Core Problem Understanding how information flows through society during disasters, identifying bottlenecks, and optimizing communication strategies to improve disaster response outcomes. ## Novel Approach: Beyond Traditional RAG Instead of conventional Retrieval-Augmented Generation (RAG), NEXUS implements: ### 1. **Semantic Propagation Graphs** ### 2. **Attention-Based Selective Broadcasting** ### 3. **Information Propagation Methods**

We built an AI system that lets officers, investigators, or trainers ask natural-language questions about any incident video traffic stops, training sessions, or field encounters and instantly get timestamped, explainable answers.

RiNI unlocks agent-to-agent context sharing, enabling separate AI agents (Work Planner and Trip Planner) to collaborate intelligently. By synchronizing professional and personal contexts, the system not only prevents conflicts for the user but also helps others in the organization schedule meetings more effectively, improving overall productivity and coordination.

Memory Engine is a metadata-aware debugging assistant that helps developers instantly recall how past errors were fixed. It combines semantic embeddings with contextual metadata, like service, environment, and version to surface the most relevant past incidents. Every fix feeds back into the system, turning your organization’s debugging history into a self-learning memory that gets smarter over time.

Context is the ultimate bottleneck of LLM capabilities. We introduce a middleware for Langchain (and soon OpenCode and others) to dynamically assemble a complete context optimized for the task at hand with every message sent to the LLM. Benchmarks (MT-RAG from IBM) show ~6% improvement in multi-step and multi-turn agent performance with a ~35% reduction in token count and cost and a ~22% reduction in latency compared to the current state of the art.

ContextNet: Real-Time Multi-Agent Collaboration Framework Traditional multi-agent AI systems are painfully slow and sequential - agents can't see what others are doing, causing duplicated work and long wait times. ContextNet solves this with MongoDB-powered real-time context sharing that enables AI agents to work in parallel. Our system demonstrates this with five agents collaborating to plan events: while ThemeAgent sets parameters, PartnerAgent, VenueAgent, JudgeAgent, and FoodAgent work simultaneously using shared context. Built on MongoDB Change Streams for instant updates and featuring a live React dashboard, ContextNet transforms how AI agents collaborate - eliminating bottlenecks, reducing costs, and providing unprecedented visibility into multi-agent decision-making for complex workflows.

A distributed AI system where specialized financial agents (Portfolio Manager, Tax Optimizer, Risk Assessor, Market Researcher, Financial Planner, and Compliance) share a hierarchical memory architecture to provide coordinated wealth management advice. The system maintains three memory types—episodic (client interactions), semantic (stable knowledge), and procedural (learned strategies)—with time-decay models and cross-agent indexing to ensure consistent, context-aware recommendations across years-long client relationships. This eliminates redundant data gathering and fragmented advice by enabling all agents to access and contribute to a unified client context stored in a shared vector database.

Multimodal Learning Memory A multimodal memory, using voyageai and mongoDB, with the ability to teach image recognition from human in the loop.

We've built a seeded RAG system that uses a lightweight knowledge cache to minimize redundant retriever calls. The LLM first tries to answer a query using cached context from previous questions; if it determines that the information is insufficient, it triggers the retriever to fetch new data and updates the cache. This approach allows semantically related or follow-up questions to be answered instantly from the existing cache, reducing latency, cost, and retrieval frequency while keeping responses contextually consistent across a conversation.

A real-time visualization tool that demonstrates how AI perception dramatically changes when given personal context/memories. Built for Problem Statement #3 (Memory-Informed Experiences), Memory Mirror extracts and stores structured memories from user input, then shows side-by-side comparisons of AI responses with and without this context. Key insight: Users can immediately see why AI fails without memory and succeeds with it. Built by a non-technical founder in 6 hours while "vibe coding" - proving accessible AI development. Features: - Real-time memory extraction and categorization - Visual memory graph with confidence scoring - Side-by-side response comparison ("The Mirror Effect") - Memory timeline and type analytics - One-click demo mode for easy testing Impact: Demonstrates the critical importance of memory in AI interactions and showcases the potential of memory-informed AI agents.

A dynamic cache that remembers what’s already been retrieved and adds only delta context when the query changes. Incremental Retrieval Cache (IRC) cuts retrieval+LLM tokens by ≥50% vs. full RAG/turn.

AI Human Journal - Document your life in a smart way The Problem: Traditional journaling is tedious and time-consuming. People struggle to maintain consistency, identify patterns in their decisions and outcomes, and extract actionable insights from their experiences. Life lessons get forgotten, and personal growth remains unmeasured. Our Solution: - Conversational Documentation: Users talk naturally with "Alex," an AI companion via video/voice, making daily reflection effortless (10x faster than typing) - Intelligent Pattern Recognition: Automatically analyzes recurring themes, decision patterns, and behavioral trends from journal entries - Semantic Memory Search: Uses MongoDB Atlas Vector Search with embeddings to find relevant past experiences instantly - ask "When did I face a similar challenge at work?" and get contextual results

Unified Capsule Memory Router (RaaS) for the Agentic Memory MongoDB hackathon: RaaS is an open-source “memory substrate” that any AI agent can call to store, retrieve, summarize, pin, and forget information with explanations and policy controls, powered by MongoDB Atlas Vector Search and Voyage AI embeddings.

HR Cortex, a multi-agent HR intelligence platform where MongoDB acts as the shared brain. Each AI agent — goals, evaluations, KPIs — writes and reads from a common MongoDB memory. With Change Streams and Vector Search, agents collaborate in real time to produce richer, context-aware employee evaluations