AGI MCP Server

Instructions

Implementation Reference

• src/memory-manager.js:93-126 (handler)

  The actual implementation of searchMemoriesBySimilarity method that performs vector similarity search using PostgreSQL's <=> operator to find memories with embeddings above a threshold, ordered by similarity distance.

  async searchMemoriesBySimilarity(queryEmbedding, limit = 10, threshold = 0.7) {
    try {
      const embeddingVector = `[${queryEmbedding.join(',')}]`;
      
      const results = await this.db
        .select({
          id: schema.memories.id,
          type: schema.memories.type,
          content: schema.memories.content,
          importance: schema.memories.importance,
          accessCount: schema.memories.accessCount,
          createdAt: schema.memories.createdAt,
          relevanceScore: schema.memories.relevanceScore,
          similarity: sql`1 - (${schema.memories.embedding} <=> ${embeddingVector}::vector)`.as('similarity')
        })
        .from(schema.memories)
        .where(
          and(
            eq(schema.memories.status, 'active'),
            sql`1 - (${schema.memories.embedding} <=> ${embeddingVector}::vector) >= ${threshold}`
          )
        )
        .orderBy(sql`${schema.memories.embedding} <=> ${embeddingVector}::vector`)
        .limit(limit);
  
      return results;
    } catch (error) {
      const truncatedEmbedding = queryEmbedding.length > 10 
        ? `[${queryEmbedding.slice(0, 5).join(',')}...${queryEmbedding.slice(-5).join(',')}] (${queryEmbedding.length} values)`
        : `[${queryEmbedding.join(',')}]`;
      console.error('Error searching memories by similarity with embedding:', truncatedEmbedding, error.message);
      throw error;
    }
  }

• mcp.js:546-552 (registration)

  The tool handler case that receives the tool request and calls memoryManager.searchMemoriesBySimilarity with the embedding, limit, and threshold parameters.

  case "search_memories_similarity":
    const similarMemories = await memoryManager.searchMemoriesBySimilarity(
      args.embedding,
      args.limit || 10,
      args.threshold || 0.7
    );
    return { content: [{ type: "text", text: JSON.stringify(similarMemories, null, 2) }] };

• mcp.js:62-86 (schema)

  The tool schema definition for search_memories_similarity including its description, input parameters (embedding array, limit integer, threshold number), and required fields.

  {
    name: "search_memories_similarity",
    description: "Search memories by vector similarity",
    inputSchema: {
      type: "object",
      properties: {
        embedding: {
          type: "array",
          items: { type: "number" },
          description: "Query embedding vector"
        },
        limit: {
          type: "integer",
          description: "Maximum number of results",
          default: 10
        },
        threshold: {
          type: "number",
          description: "Minimum similarity threshold",
          default: 0.7
        }
      },
      required: ["embedding"]
    }
  },

• src/tools/memory-tools.js:37-60 (schema)

  A standalone schema definition for search_memories_similarity in the memory-tools module, defining the same input validation structure as the main registration.

    name: "search_memories_similarity",
    description: "Search memories by vector similarity",
    inputSchema: {
      type: "object",
      properties: {
        embedding: {
          type: "array",
          items: { type: "number" },
          description: "Query embedding vector"
        },
        limit: {
          type: "integer",
          description: "Maximum number of results",
          default: 10
        },
        threshold: {
          type: "number",
          description: "Minimum similarity threshold",
          default: 0.7
        }
      },
      required: ["embedding"]
    }
  },