NKNGO

Inspiration Every autonomous AI agent wakes up stateless. A year of conversations later, it still does not know who you are, what stack you enforce, or what bugs you fixed yesterday. Long context windows (1M to 2M tokens) are an expensive, lossy scratchpad. As tokens grow, costs soar, latency spikes, and accuracy degrades due to the "Lost in the Middle" effect. While building this project with Muse Code, our agent repeatedly forgot decisions made 10 turns earlier after context compaction. We turned that exact pain into our thesis: "Agents do not have a reasoning problem. They have a state persistence problem." Tottie solves this by inverting the paradigm: applying intelligence at Write Time, not Read Time. What It Does Tottie is a self-curating agent memory layer over the Model Context Protocol (MCP) that any AI agent can mount in one configuration line. Curates at Ingestion: Evaluates turns in real time, filtering chitchat, scoring surprise, and extracting typed knowledge triples. Reconciles Contradictions: Saying "I am vegetarian now" creates a SUPERSEDE edge with temporal provenance, retiring the old preference while keeping history intact. 3-Leg Associative Recall: Combines Temporal Decay + Vector Similarity (pgvector HNSW) + Lexical Matching via Reciprocal Rank Fusion (RRF), then expands via Breadth-First Spreading Activation into a 400-token Context Card. Sleep Consolidation: Fades episodic noise and distills repeated patterns into permanent semantic rules. Living 3D Constellation Brain: A real-time WebGL visualizer (localhost:3055/?tab=graph) with living 60fps breathing pulses, electric cyan node births, and activation rays. How We Built It Meta Model API: Meta Muse Spark (muse-spark-1.3-contributor): The Observer, evaluating conversational turns in under 350ms to classify relevance and surprise. Meta Muse Spark Thinker (muse-spark-1.3-thinker): The Scribe, extracting schema-validated Subject-Predicate-Object triples, and the Consolidator during Sleep. Dogfooded with Muse Code: Continuous session watcher (dev-ref/session_watch.py) tails live Muse Code JSONL turns directly into Tottie. Storage: PostgreSQL 16 + pgvector (384-d HNSW) + FastEmbed bge-small-en-v1.5. Open Standard: Full MCP stdio server with 9 tools, 2 resources, and 1 prompt. UI: Next.js 14, React Three Fiber, Three-Forcegraph with dynamic charge repulsion. Accomplishments & Impact 22/22 Pass Rate on W5 benchmark harness with 0.0% false leakage when disabled. 91.2% Token Reduction: Serves 1,087-token cards vs 12,000+ raw history dumps. 124 Automated Tests Passing (83 backend + 41 UI tests). 100% Hackathon-Authored: All code created between Oct 27 and Nov 4, 2026.