The Hangover Part AI · Winners
What happens in the context window doesn't have to stay there. Meet the crews who built AI that doesn't forget, powered by Cognee's hybrid graph-vector memory.
Grand Winners
Two tracks, two champions. Best Use of Cognee Open Source and Best Use of Cognee Cloud.
Vinayak Sonthalia
Vinayak Sonthalia
Lethe is a self-hosted, open-source on-call incident-triage assistant built on Cognee, with one contrarian thesis: everyone builds AI memory that only grows, but the real on-call failure is memory that confidently repeats a stale fact. The dangerous moment is 3am, mid-outage, when your assistant tells you to fix a server that was decommissioned last month. It didn't malfunction — it remembered something it should have forgotten. So Lethe makes verifiable forgetting a first-class product feature. What it does: You feed it your team's runbooks and post-mortems as plain English — no schema, no tags. Cognee's cognify() builds a knowledge graph (Kùzu) plus a vector index (LanceDB) from the prose automatically. Ask a real incident question and Lethe answers in runbook style — naming exact systems and actions, with every claim cited to its source document. When a system is decommissioned, Lethe performs a hard delete — raw files, graph nodes, edges, and vector embeddings — and prints a measured receipt ("2 documents · 10 graph nodes · 18 relationships · 0 chunks left in the index") with a live re-query proof and a downloadable Certificate of Erasure. Ask the exact same question again and the answer flips to the systems that still exist; ask about the dead one and it answers "not documented in the runbooks" instead of hallucinating. Beyond the hero feature: a two-layer forgetting model (hard delete + reversible demote built on Cognee's feedback weights), a human-gated curation loop (deterministic memory-health score at 0 tokens; aging knowledge sinks automatically; hard deletes always queue for human approval), an audit timeline of every learn/review/demote/forget event, isolated workspaces, bring-your-own-key (any OpenAI-compatible provider or fully local via Ollama; embeddings run on local fastembed at $0), an MCP server so Claude Code/Cursor can triage and decommission directly, and 30 chapters of in-app documentation (/learn) explaining every design decision. The proof: I benchmarked Cognee's four advertised capabilities against plain RAG at temperature 0 — three tied with the baseline; only forget clearly won, so I built the product around the one honest, provable win. Forgetting itself is measured with a blind judge (a different model family scores every answer 0–2), replicated twice with different model pairs: stale advice 1.0→2.0, abstention/honesty 0→1.0, unrelated answers unharmed (2.0→2.0), and every decommissioned system provably wiped from the index. Per-question JSON is in the repo. Built solo in one week: product, landing page, benchmark suite, MCP server, docs, and demo video.
Rajdeep Singh
Rajdeep Singh
Classroom Memory is a Cognee Cloud-powered classroom memory system. The idea is simple: most learning apps remember content, but they do not really remember students. They can show quizzes, marks, and dashboards, but they usually cannot answer questions like: what does this student already know, what are they forgetting, what are they ready to learn next, and what should the teacher teach the whole class tomorrow? Classroom Memory treats a classroom as a memory problem. Each student gets their own private learning memory. The student page shows an adaptive quiz, a live prerequisite mastery graph, a personal ask-memory box, a learning timeline, and a progress report. Every quiz answer updates that student’s mastery state and reshapes the graph. The teacher page reads across the class and shows a class heat map, a graph-reasoned teaching plan, class-level recall, and targeted review assignment. So instead of only saying “the class is weak in recursion,” the app can reason over prerequisites and say something more useful, like: “Teach functions first, because students are not ready for recursion yet, and functions unlock many downstream concepts.” The project is built as a working product prototype, not only a demo. It includes student enrollment, class roster setup, curriculum import from JSON, SQLite persistence for exact mastery state, real Cognee Cloud memory calls, and a verification script that proves the live Cognee Cloud lifecycle.
Best Blog
The write-up that best told the tale of building with Cognee memory, worth a Keychron mechanical keyboard.
Mahimai Raja J
⌨️ Keychron KeyboardAuthor of the winning blog — a solo builder who shipped RealtyRecall, an always-on voice receptionist that gives real-estate agents a memory of every caller, and wrote the standout end-to-end account of doing it with Cognee.
RealtyRecall is an always-on voice receptionist for solo real-estate agents. Solo agents miss calls all day showings, driving, meetings and every missed call is a lost lead. RealtyRecall answers every call in the realtor's own voice and persona, qualifies the buyer, recommends only the agent's real connected listings, books showings on their calendar, and texts them the lead. What sets it apart is memory: it's built on a Cognee knowledge graph, so it recognizes returning callers by phone, recalls what they were looking for, and proactively surfaces newer homes near what they liked. Realtors onboard in one click by pasting their website URL. RealtyRecall crawls it into a per-tenant knowledge graph. It's fully multi-tenant (each agency isolated via Cognee NodeSets), deployed live, and includes a console that visualizes the memory graph growing in real time as calls happen.