mcp-titan — Cursor Rules

• init_model `typescript interface InitModelParams { inputDim?: number; // Default: 768 hiddenDim?: number; // Default: 512 memoryDim?: number; // Default: 1024 transformerLayers?: number; // Default: 6 numHeads?: number; // Default: 8 ffDimension?: number; // Default: 2048 dropoutRate?: number; // Default: 0.1 maxSequenceLength?: number; // Default: 512 memorySlots?: number; // Default: 5000 similarityThreshold?: number; // Default: 0.65 surpriseDecay?: number; // Default: 0.9 pruningInterval?: number; // Default: 1000 gradientClip?: number; // Default: 1.0 } const response = await callTool("init_model", { inputDim: 768, memorySlots: 10000, transformerLayers: 8, }); ` • forward_pass `typescript interface ForwardPassParams { x: number[] | string; // Input vector or text memoryState: { shortTerm: number[]; longTerm: number[]; meta: number[]; timestamps: number[]; accessCounts: number[]; surpriseHistory: number[]; }; } const { predicted, memoryUpdate } = await callTool("forward_pass", { x: "const x = 5;", // or vector: [0.1, 0.2, ...] memoryState: currentMemory, }); ` • train_step `typescript interface TrainStepParams { x_t: number[] | string; // Current input x_next: number[] | string; // Next input } const result = await callTool("train_step", { x_t: "function hello() {", x_next: " console.log('world');", }); ` • get_memory_state `typescript interface GetMemoryStateParams { type?: string; // Optional memory type filter } const state = await callTool("get_memory_state", {}); // Returns: { stats: { meanActivation: number; patternDiversity: number; surpriseScore: number; timestamps: number[]; accessCounts: number[]; }, capacity: number; status: "active" | "pruning" | "error"; } ` • manifold_step `typescript interface ManifoldStepParams { base: number[]; // Base memory state velocity: number[]; // Update direction } const newBase = await callTool("manifold_step", { base: memory.meta, velocity: [0.1, 0.2, ...] // or analysis.stats.meanActivation }); ` • prune_memory `typescript interface PruneMemoryParams { threshold: number; // Pruning threshold (0-1) } await callTool("prune_memory", { threshold: 0.65, }); ` • save_checkpoint `typescript interface SaveCheckpointParams { path: string; // Checkpoint file path } await callTool("save_checkpoint", { path: "/backups/memory-snapshot.json", }); ` • load_checkpoint `typescript interface LoadCheckpointParams { path: string; // Checkpoint file path } const recovered = await callTool("load_checkpoint", { path: "/backups/last-stable.json", }); ` • reset_gradients `typescript // No parameters needed await callTool("reset_gradients", {}); `

• Safe Tensor Operations `typescript const result = memoryManager.wrapWithMemoryManagement(() => { const input = vectorProcessor.processInput(userInput); const validated = vectorProcessor.validateAndNormalize(input, [768]); return model.forward(validated, currentState); }); ` • Async Memory Operations `typescript const result = await memoryManager.wrapWithMemoryManagementAsync(async () => { const encoded = await vectorProcessor.encodeText(text); const processed = vectorProcessor.processInput(encoded); return model.forward(processed, currentState); }); ` • Secure State Saving `typescript const encryptedState = memoryManager.wrapWithMemoryManagement(() => { const state = model.getMemorySnapshot(); return { shortTerm: memoryManager.encryptTensor(state.shortTerm), longTerm: memoryManager.encryptTensor(state.longTerm), meta: memoryManager.encryptTensor(state.meta), }; }); `

• Memory-Safe Code Analysis `typescript const result = await memoryManager.wrapWithMemoryManagementAsync(async () => { // Process code input const processed = vectorProcessor.processInput(codeBlock); // Forward pass with memory management const { predicted, memoryUpdate } = await model.forward( processed, currentMemory ); // Check memory health const state = await model.getMemoryState(); if (state.stats.surpriseScore > 0.85) { // Update memory safely const newBase = await model.manifoldStep( memory.meta, state.stats.meanActivation ); await model.trainStep(processed, predicted); } return predicted; }); ` • Secure Memory Maintenance `typescript const maintenance = await memoryManager.wrapWithMemoryManagementAsync( async () => { // Save current state securely const encrypted = memoryManager.encryptTensor(currentState); await saveToFile(encrypted); // Prune memory if needed const health = await model.getMemoryState(); if (health.capacity < 0.3) { await model.pruneMemory(currentState, 0.5); } // Verify health const postPrune = await model.getMemoryState(); return postPrune.status === "active"; } ); `

• Initialization Requirements • MUST call init_model at start of session • MUST specify at least inputDim and memorySlots • MUST handle initialization errors • Forward Pass Requirements • MUST call before processing any input • MUST provide complete memory state • MUST handle both vector and text inputs • Training Requirements • MUST call after every interaction • MUST provide sequential inputs • MUST maintain context continuity • Memory State Requirements • MUST check every 5 interactions • MUST monitor all stats fields • MUST act on critical thresholds • Maintenance Requirements • MUST prune when capacity < 30% • MUST save checkpoints every 5 minutes • MUST reset on gradient explosions